Gnuspeech Monet Manual 08 24

User Manual:

Open the PDF directly: View PDF PDF.
Page Count: 52

Gnuspeech Monet Manual 0.9
Gnuspeech: the speech synthesis database creation,
modification, and articulatory synthesis software suite
David R. Hill, University of Calgary
Based on work over more than three decades by the author, Leonard Manzara, Craig Schock,
Ian Witten, Wiktor Jassem, Steve Nygard, Dalmazio Brisinda, Marcelo Matuda, and many
talented undergraduate students (See “Thanks to those who have helped”
http://www.gnu.org/software/gnuspeech, accessed 2015-07-24). Thanks also to Richard
Stallman and the Gnuisances, Karl Berry and Brandon Invergo at the Free Software
Foundation for their enlightened support.
i
Copyright ©2002, 20012, 2015 David R. Hill
This manual is for Gnuspeech version 0.9, 23 August 2015
Permission is granted to copy, distribute and/or modify this document under the terms
of the GNU Free Documentation License, Version 1.3 or any later version published by
the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no
Back-Cover Texts.
A copy of the license is available at: http://www.gnu.org/licenses/fdl-1.3.en.html—the
Free Software Foundation site (accessed 2015-08-15)
ii
Summary
Monet1comprises three main divisions serving distinct functions: (1) a Speech Server
that translates ordinary text into spoken output, via a specialised phonetic script with
automatically added rhythm and intonation information; (2) a speech synthesis database
editor and manager with a Graphical User Interface (GUI); and (3) a GUI front end to
the Speech Server that is intended to allow databases to be tested as they are developed,
but also allows the speech synthesis to be evaluated, including the effects of intonation,
and can provide carefully managed test utterances for subjective testing.
These divisions form the core of the Gnuspeech software, providing: (1) real-time ar-
ticulatory speech synthesis; and (2) providing databases creation facilities for arbitrary
languages, research facilities for linguistics research, speech therapy, language-learning,
and other speech-research-related topics. A platform-independent synthesis engine—
GnuspeechSA—provides command-line access to the speech synthesis capability (See Ap-
pendix D).
Monet was one of two primary tools used to create the original databases for the
spoken English that can now be produced. The other tool was an interactive application,
currently known as TRAcT 2Two other important helper tools are also involved in the
process of creating new databases: a basic editor to compile the pronouncing dictionary
that is used (Monet has a dictionary of over 70,000 words, plus a means of constructing
derivatives); and a means of generating detailed sound spectrograms of utterances to
validate the correctness of both static and dynamic sounds being developed using Monet.
In connection with the dictionary, a letter-to-sound converter is highly desirable to catch
any words that cannot be found or constructed using the dictionary.
The original complete “TextToSpeech” kits that represented the NeXT Computer
implementation also included auxiliary applets, most notably “BigMouth” that used the
Speech Server to speak files, speak scratch-pad text, and vary some speaker characteristics,
and “PrEditor”, that can be used to create dictionaries specific to a particular user,
or a particular application. These are not yet ported and will be the subject of later
releases. The functions associated with database creation are not completely ported, but
are described as if complete to guide those who continue to work on the port and further
development.
This manual provides a description of the structure and use of Monet together with
selected tables of the data used by the system. A quick introduction to GnuspeechSA on
the Macintosh is also provided.
1“My Own Nifty Editing Tool”, designed and implemented by Craig Schock as the system architect
for the original NeXT computer version, to provide the foundation tool for creating “event-based” speech
data based on earlier and ongoing research by the author. Both the creation of the system, and its use
to develop the articulatory speech databases, involved sufficient artistry that it seems very appropriate
to maintain the artistic name connection. The current system is mostly ported to both GNU/Linux and
the Macintosh. The implementation of Gnuspeech under GNUStep is reasonably well documented by
the original NeXT manual that also accompanies this release, though it needs to be updated (real soon
now!).
2TRAcT—Tube Resonance Access Tool—is available as part of this GNU project release, and includes
a manual. TRAcT allows a user to experiment with static configurations of the vocal tract model that
provides the low-level articulatory-synthesis waveform-output from the system (and from standalone use
of the Speech Server), and was necessary in developing the configurations, or “postures” whose dynamic
variation produce speech.
iii
[This page is intentionally blank]
iv
Contents
1 Purpose of the system 1
2 System capabilities 1
3 Introduction 3
4 System overview and rationale 7
4.1 Introduction...................................... 7
4.2 Main components, subsystems, and databases . . . . . . . . . . . . . . . . . 8
4.2.1 Monet..................................... 8
4.2.2 The Gnuspeech database that is edited using Monet . . . . . . . . . 10
5 Monet-based Speech Server access 11
5.1 Overview........................................ 11
5.2 Segmentalsynthesis ................................. 11
5.3 IntonationandRhythm............................... 13
5.3.1 The intonation parameters & the intonation window . . . . . . . . . 15
6 Database creation and modification 18
6.1 Preamble........................................ 18
6.2 Gettingstarted.................................... 18
6.3 Creating/editing the database . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
6.3.1 Categories, Parameters, and Basic Timing Framework . . . . . . . . 19
6.3.2 PostureEditor................................ 20
6.3.3 Prototypes, Equations, and Transitions . . . . . . . . . . . . . . . . . 21
6.4 Prototype Manager and Transitions . . . . . . . . . . . . . . . . . . . . . . . . 22
6.4.1 Equations................................... 22
6.4.2 Transitions .................................. 22
6.4.3 Prototype Manager: Special Transitions . . . . . . . . . . . . . . . . . 26
6.5 TheRuleManager.................................. 26
6.5.1 Introduction ................................. 26
6.5.2 Expressions.................................. 27
6.5.3 Symbols.................................... 28
6.5.4 Parameters.................................. 28
6.5.5 SpecialParameters ............................. 29
6.5.6 Meta Parameters and Comments . . . . . . . . . . . . . . . . . . . . . 30
6.5.7 TheRuleTester ............................... 30
7 References 31
A Some Monet posture parameter and phonetic data 34
B Pronunciation Guide 35
B.1 Font comparison: Webster, Gnuspeech (Trillium), and IPA . . . . . . . . . . 35
B.2 Illustration of some pitfalls between American and British English . . . . . 38
B.3 Syllables, stress, and American versus British English . . . . . . . . . . . . . 39
v
C Rewrite rules for Monet 40
C.1 Idiosyncratic rewrite rules and allophonic requirements . . . . . . . . . . . . 40
C.2 Linking /r/, intrusive /r/ and glottal stop insertion . . . . . . . . . . . . . . 40
D GnuspeechSA Speech Server command line tool 41
D.1 Introduction...................................... 41
D.2 Installing GnuspeechSA on the Macintosh . . . . . . . . . . . . . . . . . . . . 41
E Outline of Halliday’s British English intonation notation 42
E.1 British English Intonation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
vi
1 Purpose of the system
Monet is a special purpose interactive GUI-based editing system built to facilitate the
creation and modification of databases that allow computer synthesis of spoken language.
Monet has a recursive flavour—in order to carry out its function it must include a means
of speaking the rudiments language it is developing, achieved by ensuring “speech” can
be produced even before any data, apart from one posture, has been entered.
2 System capabilities
The current version of Monet, shown in Figure 1 includes:
1. The real-time “Speech Server”3that converts text to phonetic descriptions, to syn-
thesiser parameters, and thence into a speech waveform. The Speech Server com-
prises:
(a) a database containing posture data (including, but not limited to: tube radii;
marked and unmarked durations; rules for combining n-phones; transition
shapes used by the rules; and special transitions for things like noise-bursts
and micro-intonation);
(b) a parsing module for pre-processing ordinary text to account for the conven-
tions specific to the target language (such as how dates and acronyms should
be pronounced);
(c) a pronouncing dictionary with methods for creating derived words;
(d) a language-specific set of letter-to-sound rules to convert words for which there
is no dictionary entry or derived word;
(e) a waveguide (“tube”) model of the human vocal apparatus to perform the
conversion of parameters into a speech waveform for output;
(f) a computational model of rhythm, applied to the phonetics, based on natural
speech data derived from the specific target language—English in this case;
and
(g) a computational model of intonation—the rise and fall of pitch—based on
language-specific linguistic models available in the literature and refined/tested
by subjective experiment. Such a model is actually not practical in early
development of a language database and we used monotone in the English
development and added the intonation module later, though some simple pitch
movement allowed a better assessment of results.
2. GUI-based access to the Speech Server, with informative displays and controls:
(a) a “Synthesis Window” that allows entry of text or posture strings that are to
be spoken, as well as allowing the thirty-two articulatory control parameters
(sixteen regular and sixteen special) to be plotted and displayed, along with
timing and posture-type information. The speech events can be displayed, the
graph images saved, and the input synthesised to provide speech output, or
3the Speech Server is also the core of the cross-platform applications related to speech output developed
by Marcelo Matuda that are part of this release. (See Appendix D
1
sent to a sound file. It uses the Speech Server for synthesis and is essential for
checking the developing database when developing a new one, or modifying an
existing one, as well as allowing the creation of stimuli for subjective testing;
(b) a “Synthesizer Control Panel” that allows the “Control Rate” parameters (such
as vocal tract–tube–length, and nasal passage radii) to be set;
(c) an “Intonation Window” that provides a graph of the intonation contour, show-
ing the postures and timing. It includes the ability to modify the contour;
(d) an “Intonation Settings” window providing control of the intonation contour
parameters (such as component ranges, and whether smoothing is applied) and
allows the speech “Tempo” to be changed.
3. A special-purposes GUI-based editor—not yet fully ported—that provides the abil-
ity to create and edit the databases used by the Speech Server. It includes:
(a) entry of posture names, parameter values, timing, and categories;
(b) means for creating prototypes for parameter transitions, based on arbitrary
timing points with chosen descriptive names. The transitions are not shapes
per se, but are graphs showing the percentage change from one target to the
next, plotted against time, and showing the designated timing points. Two
kinds of transition prototypes are provided, regular and special. A Prototype
Manager provides a higher-level means of managing these;
(c) a rule manager that specifies which transition profiles should be applied to
which combinations of postures for which parameters. If postures are thought
of as phones, provision is made for n-phone combinations, where nvaries from
2 to 4 (an isolated phone still requires transitions from and to silence);
(d) a rule tester to make it possible to see which rules will apply to a given string
of postures as they are entered into the tester and shifted.
Figure 1: Overview of Monet, showing components ported so far
2
The TRAcT manual that is part of this release provides background on the origins
of speech synthesis and the progress that has been made, as well as the technical details
of the waveguide vocal tract model or “Tube Resonance Model” (TRM). The TRM is
properly thought of as a low-level synthesiser as opposed to the complete system which
is a high-level synthesiser. The low level synthesiser is akin to a trumpet—it only plays
music when operated by a trained trumpet player (the player is analogous to the high-
level control that produces speech from the low-level synthesiser). Papers describing the
original research used to establish the databases, the rhythm model, and the intonation
model, are cited.The most recent paper (Hill 2015) gives more detail on the background,
and outlines the procedure by which the databases were created. The work is all original,
drawing on long-term research in many laboratories, other university departments, and
by team members in the author’s laboratory, as well as members of the Trillium Sound
Research team. Trillium Sound Research was a University of Calgary technology spin-off
company that created the original NeXT implementation of the system (then known as
various levels of the “TextToSpeech Kit”). The company was dissolved, and the software
donated to the Free Software Foundation (FSF) under a General Public Licence and Gen-
eral Document Licence—hence the GNU Project “Gnuspeech”. Acknowledgements ap-
pear, amongst other places, on the http://www.gnu.org/software/gnuspeech—Gnuspeech
home page which is on the FSF site ( accessed 2015-06-05).
3 Introduction
Unrestricted speech synthesis-by-rules efforts have often used methods developed in the
50s and 60s to simulate speech production by feeding information about spectral features
to a source-filter model of the vocal apparatus. The devices comprise a pulsed energy
source, and a set of bandpass filters that imitate some of the excitation and resonant prop-
erties of the oral and sometimes nasal passages in the human head [e.g. Lawrence (1953);
Fant (1956)]—“low-level synthesis”. As noted, “high-level synthesis” is then needed to
provide the data and algorithms required to make the low-level synthesiser speak a partic-
ular language. Early work by other researchers provided the data and methods (high-level
synthesis) needed to drive these models to produce synthetic speech—for example Liber-
man, et al. (1959); Holmes et al. (1964). The overall approach has been given a variety
of names, but “formant synthesis” seems the most descriptive, since the variable driving
data mainly comprise the centre frequencies of the resonances of the vocal tract4inter-
acting with the energy input from the vocal folds, plus the various noises produced as a
result of the passage of air through constrictions in the tract, including the vocal folds.
“DECtalk”—derived from Allen’s (1987) “MITalk”—is a formant synthesis approach.
Formant synthesis is widely used—for example, by the distinguished physicist Stephen
Hawking to allow him to “speak” despite his ALS, which prevents him from speaking
naturally. Other systems are closely related. More recently “concatenative synthesis” has
become popular. In this approach real speech is captured by some suitably sophisticated
method (typically as Linear Predictive Coding—LPC, which allows easy separation of
source excitation from the filtering effect of the vocal tract). The speech is segmented to
collect suitable basic elements, and these are then rearranged and concatenated to produce
new utterances. Although the speech is recognisably natural in sound quality, it suffers
from: limitations inherent in the process which reduce the intelligibility: lack of natural
4The resonances correlate with the frequency peaks or “formants” in the output spectrum.
3
intonation and rhythm; and difficulty in producing voices that have their own character
distinct from the original recordings. Thus both formant and concatenation methods still
suffer from restrictions of various kinds that interfere with the potential for naturalness
in unrestricted speech—though, for restricted purposes, concatenated natural speech can
be very effective. Waveform concatentation is the principle method underlying the Fes-
tival system at Edinburgh University’s Centre for Speech Technology Research (CSTR
undated). This project modularises the synthesis process so that researchers can work on
a single module (say intonation management) without having to create and manage the
entire synthesis process.
Mark Tatham’s (2000) SPRUCE project at the University of Essex (UK) has been
described as a system that provides the high level synthesis needed to drive both formant
synthesis and concatenative synthesis. The emphasis appears to be on concatenative
synthesis.
In 1993/4, building on fundamental work by Carr´e, Fant and others (e.g. Fant & Pauli
1974; Carr´e & Mrayati 1994), Cook (1989, 1991), Smith (1987a, 1987b) the author’s team
developed a speech synthesis system that uses a wave-guide approximation to the real
vocal tract oral/nasal filter as the filter component, as opposed to individual filters for
each formant, thereby providing an articulatory model—also called a “tube model”, “tube
resonance model” (TRM), or “waveguide model” ( Hill et al. (1995),). The work on the
waveguide model at the Center for Computer Research in Music and Acoustics Stanford,
by Julius Smith, Perry Cook and their colleagues, was a crucial element in the original
NeXT implementation of their waveguide model, and the basic design has carried through
into the current system. Such a model emulates5rather than simulates the resonant
behaviour of the vocal tract because the tube model behaviour maps directly onto the
articulatory and acoustic characteristics of the real vocal tract and nasal passage tube
structures, rather than simply reproducing a partial output spectrum by some technical
contrivance. For example, higher formants are produced naturally, and they vary, rather
than having to be simulated by some form of fixed lumped-filter approximation. Also the
nasal-oral energy balance is implemented directly, oral and nasal radiation are modelled
directly, and so on.
Figure 2: DRM regions 1 through 8 shown
nominally on a sagittal section of the head
The previous barriers to using a tube
model for speech were three-fold. First
there was the problem of controlling the
many sections—typically 40—required for
the tube approximation, in real-time, with-
out instability or self-defeating approxima-
tions; secondly there was the problem of
providing the complete database that rep-
resented the behaviour of many parts of a
real vocal apparatus6speaking a particular
language; and thirdly, until the digital age,
there was a problem with the stability of
the analogue circuits that implemented the
model. The work of Fant, Carr´e and their
colleagues (op. cit.) provided the theoreti-
5Emulate: to equal or surpass, especially by imitation
6Lungs, trachea, vocal folds, oral tube, tongue, teeth, cheeks, velum, nasal tube, lips, mouth orifice,
and nostrils.
4
cal basis for solving the control problem for the tube model. Based on a formant sensitivity
analysis research by Fant and his colleagues, Carr´e and Mrayati devised what they call
the Distinctive Region Model (DRM) that provides a reasonably accurate model of artic-
ulation, that is related to the known properties of the real vocal tract, but requires only
eight independently controlled sections instead of the forty or so that would be needed
if the constraints of the vocal tract of speech were ignored. The topic is discussed more
fully in the paper by Hill, Manzara & Schock (1995) and in the manual for TRAcT that is
part of this release. TRAcT is concerned with the TRM and its manipulation. The DRM
control sections applied to the TRM correspond closely to the distribution of articulatory
possibilities in the real vocal tract so that, even though the traditional parameters such
as jaw rotation, tongue height, lip opening, and so on are not used directly, the model is
truly an articulatory model, and the traditional parameters could be used at a meta-level
to define the changes in the DRM regions. Provision for this intended extension has been
made in the basic framework of the system.
A description of how the English database was built, what it contains, and why it has
the form it has, appears in Hill (submitted).
In addition to the articulatory parameters that model the changes in the vocal appa-
ratus associated with the succession of speech sounds Monet also models British English
pitch contours to control the variation in pitch (intonation) over an utterance, as well as
a rhythm model, based on extensive research in the author’s university laboratory dealing
with relative and absolute time duration of postures. It is the variation in posture du-
rations that determines the rhythmic details of the speech that—in conjunction with the
intonation—are so essential for excellent intelligibility. The rhythm model is explained in
Hill, Jassem & Witten (1979) and Jassem, Hill & Witten (1984). The intonation model is
broadly based on Halliday’s model of intonation in spoken English (Halliday 1970), and
is integrated with the rhythm model, as it must be. Indeed, it is not possible to describe
the Halliday intonation model without also specifying a well-defined rhythmic structure.
In dealing with the machine perception and production of speech, a number of techni-
cal terms must inevitably be used in order to achieve precision of expression. The reader’s
attention is drawn particularly to the terms associated with speech sounds (phones,
phonemes, postures, etc) and the basic concepts associated with rhythm and intonation.
A “conceptionary” for speech and hearing in the context of machines and experimentation
(Hill 1991) provides a convenient source of such conceptual knowledge.
Especially note that “phonemes” are not speech sounds. They are categories of speech
sounds. Sounds fall in the same phoneme category for a given language if the difference
between them does not distinguish between any two words in that language. Thus the
sounds in a given phoneme category—called “allophones”—may be quite varied acousti-
cally, and may result from a variety of quite different articulatory causes. Sounds in the
same phoneme category may be as different acoustically as upper and lower case letters
are different visually. Consider the acoustic realisation of the spoken English /r/ phoneme
across various dialects—educated Southern English, General American, and Scottish, for
example (Pullum & Ladusaw 1986: p 131). Equally, allophones from different phoneme
categories may be rather similar acoustically. For example, the actual sounds (phones)
produced as different vowel phonemes may overlap for different speakers and phonetic
contexts (Ladefoged & Broadbent 1957). This is why we prefer to work from the concrete
basis of speech postures. Speech postures can easily be related to the phoneme categories
of a language, but they are not phonemes. A series of postures articulated in succession
by Monet, will produce a series of phones—instantiations of phonemes. The term phone
5
is sometimes used interchangeably with posture, but the postures in any series interact
with and modify each other, which is why the phones representing the same phoneme in
different phonetic contexts are different—they are technically called “allophones”. Thus,
especially for an articulatory speech synthesiser, the postures and associated interpolation
rules, plus special events, time quantities, and intonation contours (the latter—following
Halliday—are called tones), are the truly basic entities. The notation /r/ represents the
phoneme for the “r” sound in any English dialect while [r] represents a particular allo-
phone of the /r/ phoneme. These are called broad and narrow transcriptions respectively
when the notation may used by phoneticians to transcribe utterances. The broad tran-
scription is a very high-level transcription that assumes an understanding of a particular
dialect to fill in the missing details that describe the sounds accurately. The narrow tran-
scription uses all kinds of additional annotations, called diacritical marks, to indicate the
exact properties of the sounds in a given utterance context. Thus a broad transcription
is phonemic while a narrow transcription is phonetic and describes the individual allo-
phones explicitly. The full gory details of this topic may be pursued in any decent text
on phonetics or phonology.
The original “Version 1.0” of Monet that still runs on the NeXT computer is complete
in the sense that it has been successfully used to create a complete articulatory database
for spoken English, including rhythm and intonation. However, it could well do with
additional productivity-enhancing features such as better links between various data views
and editing tools. The usability could be improved, and all components are the subject
of ongoing research and improvement in the ported version. Some improvements of this
kind have already been made in this current version. However, it is now designated 0.9
because the database creation components have yet to be fully ported, though the interface
widgets are all in place with much of the driver code, but some final details have not been
incorporated.7Whilst all the components for producing synthetic speech are there, as
diagrammed in Figure 1, and can be used for evaluation and simple experiments, the yet-
to-be-fully-completed database creation elements cannot yet be used to create another
language, or edit the existing English databases. The heart of Gnuspeech—the kernel-
based Speech Server, is available as an operating system service to allow it to be used as a
Service on the Macintosh, or to be incorporated in new applications written by developers,
but auxiliary applets such as “BigMouth” (which provides the ability to speak files, or
text from a scratch pad) or “PrEditor” (which provides the ability to enter words and
pronunciations into the “User” and “Application” dictionaries), are not yet ported. The
main dictionary can easily be edited using an editor such as Aquamacs (the Macintosh
version of Emacs—a free download). GnuspeechSA incorporates the Speech Server for
platform-independent use (see Appendix D).
As noted above, Monet was one of the tools used to create the databases associated
with Trillium’s unique TextToSpeech system based on articulatory synthesis. The other
components used in that work included: (1) Trillium’s interactive Tube Resonance Model
access tool “Synthesizer” application8; together with (2) spectrographic analysis and dis-
play tools—specifically a Kay “Sonagraf”—at the time,9dictionaries, and so on. The
7There is also the original NeXT version manual which is a rough guide for the GNU/Linux port,
since the full port runs under GNUStep, itself a port of NEXTStep/OpenStep.
8Now renamed TRAcT for “Tube Resonance Access Tool”; the name “Synthesizer” was misleading.
9There is now an excellent software tool—Praat (Boersma 2001; van Lieshout 2003) that: (a) provides
high quality spectrographic analysis by computer; (b) is available for no payment; and (c) renders the
Kay Sonagraf unnecessary.
6
complete system is suitable for:
•speech output for end users;
•incorporation of speech output into applications by software developers;
•use in speech research by university and industry research laboratories; and
•further development of speech synthesis methods and additional languages.
4 System overview and rationale
4.1 Introduction
Monet is organised around a time/parameter-value framework that assumes speech is
formed by a vocal apparatus moving successively from one vocal posture to another,
subject to articulatory constraints and involving contextual influences. Postures can affect
neighbouring postures even when they are not immediately adjacent. Silence is a posture,
just as much as any vowel or consonant posture, and its specification may equally depend
on the context. Different versions of silence seem to be required—for example, rest versus
a glottal stop. Silence postures are assigned a posture symbol that includes an initial “q”,
except for the “rest” posture,“#” and standard silence “∧”.
Postures are often loosely referred to as phones which is not strictly correct. The term
“posture” describes “that which generates a sound”, rather than the sound itself—a phone.
Postures are not strictly equivalent to phonemes either. As already discussed, phonemes
are categories of sound whose realisations vary according to their specific phonetic context
and other factors. In fact postures in continuous speech are related to phonemes in
much the same way that phones are. Postures in continuous speech are specific context-
dependent instantiations that generate allophones as their associated sound output.
Monet assumes that speech is to be produced by a speech synthesiser that is fed varying
parameters at some time rate. In the original development, no assumptions were made
about the nature of the synthesiser that would be supplied with synthesis data, except for
the assumption that there is a special parameter controlling pitch variation that should
be manipulated specially in order to provide intonation contours. The event-based time
framework for composing the parameter generation has proved most appropriate for an
articulatory low-level synthesiser, such as the TRM.
It is assumed that each posture (corresponding to a vocal tract configuration) can be
defined in terms of the parameters used to control the synthesiser, but that the parameter
values so defined are not necessarily the same for all instantiations (realisations) of the
posture—they usually vary with context and other factors; nor do the parameters nec-
essarily all take on their characteristic values for the posture at the same time—the lips
and tongue move independently for an articulatory synthesizer. Each parameter for each
posture moving to another posture can have an individually specified transition shape,
timing, and percentage of the target value, if needed. The parameter transitions are
created within Monet and referenced for each posture combination (digram, trigram and
tetragram sequences), according to the rule that applies (also created within Monet),
and chosen to control the compilation of the parameters for the specific sequence of pos-
tures. The compilation rules are ordered and accessed from the most specific to the most
7
general—the final rule encountered being simply phone>>phone, which is always present
from the beginning, and is the default when no other rule applies.
The basic strategy is to always have a default available for any required synthesis
detail so that something will be generated with even minimal database information.This
facilitates getting started with database creation.
The time framework is constructed starting with a basic framework (“Major Event
Times” and “Posture Value Targets”). This framework is based on fixed posture targets
occurring at fixed times, but the system then provides mechanisms for specifying depar-
tures from the framework in a very flexible and comprehensive manner. In particular,
although the underlying time framework exists, the main governing principle for time
rests on the occurrence of speech events—times at which events, such as a parameter
starting to change, happen, for which named event times are specified. The target-time
framework is simply a foundation for building the more complex reality. The view is re-
lated to research on muscle action groups due to Condon & Ogston (1970) and the more
recent work by Richardson et al. (2008).
Because we have used Monet—the Graphical Database Editor—exclusively for working
with our tube-model-based articulatory synthesiser to create the databases for English
speech synthesis, the remainder of this document will assume such an arrangement, in
order to provide concrete examples when discussing the system. An account of the TRM
approach to synthesis appears in Hill, Manzara & Schock (1995) already cited, and the
GUI interface application TRAcT—the “low-level” synthesiser, which is documented in
the TRAcT manual accompanying this release.
4.2 Main components, subsystems, and databases
It should be emphasised that Gnuspeech already includes a complete database for spoken
English as required for the Speech Server. It is not necessary to create new databases, or
to modify existing ones, in order to produce articulated English speech output (including
intonation and rhythm). In fact, although some Monet components (those that com-
prise the Speech Server) provide the “brain” (the “high-level” synthesis component) that
translates input text into the parameters to drive the synthesiser, the end user interested
only in the existing speech output capability of GnuSpeech will not need to have any
understanding of Monet at all, other than how to use it to speak, or—if an application
developer—how to incorporate the Speech Server in an application—currently most easily
done by using GnuspeechSA (Matuda 2015) and see Appendix D. However, Gnuspeech
Monet goes far beyond providing an articulatory text-to-speech output means.
4.2.1 Monet
The Graphical Database Editor “Monet” provides facilities that allow for:
•Posture-specific data entry
–“Data Entry”: defining posture attributes
∗posture category names (e.g. “contoid”, “stopped”);
∗parameters names, ranges and default values (e.g. for “microint”, “r1”);
∗meta-parameter names, ranges and defaults (not currently used); and
∗“formula symbol” ranges and default values for use in formulae;
8
–“Postures”: entering posture data;
∗adding or subtracting individual postures names;
∗specifying the categories to which a given posture belongs;
∗entering parameter data applying to specific postures;
∗entering specific posture meta-parameter data (currently not used); and
∗defining the formula symbol values applying to specific postures;
•Posture-sequence parameter transition definition and management
–“Prototype Manager”: managing the “Equation”, “Transition”, and “Special
Transition” prototypes
∗naming, forming, and editing the equations governing the event timing of
transition (e.g. “TriphoneDefault”, “endOfStopClosureOnset”); the equa-
tions are arranged in groups (named as needed) and are formed in terms
of the timing symbols set up for the postures that are involved in the com-
putation; the defined symbols may be used in the computation of further
symbols, or in the specification of points in the transition interpolation
graphs (see below); they could also be used to modify target values;
–“Transition Builder” ;
∗Naming and graphically creating/editing “Parameter” interpolation graphs
(e.g. “newDiphoneDefault”, “bLipClosure”); the graph values represent
the percentage of the change between parameter targets occurring as time
progresses. Thus a rising graph can lead to a falling parameter value.
–“Special Transition Builder”;
∗Naming and graphically creating/editing “Special Parameter” interpola-
tion graphs (e.g. “kAspBurst”, “vlessStopMicroInt”); the graphs represent
changing absolute values of the particular parameter occurring as time pro-
gresses; the absolute values are combined with the regular transition values
by superposition.
Note: Both transition types use timing derived from basic time and target def-
initions according to formulae that may be defined arbitrarily as noted above.
Different symbols (•,▲, and ∎) for points are chosen depending on whether
they apply to the first, second or third posture transition region (the latter
occurring in a tertraphone posture combination). A “phantom point” is used
as placeholder where two transition profiles overlap. The purpose of a phantom
point is for display purposes only, and it means that the actual value of the
point comes from the abutting profile.10 The synthesiser parameter prototypes
for the different parameters needed to create a given n-phone can all be differ-
ent, and will implicitly define deviations in timing and target values, as well
as the actual shape of the movements required. Which prototypes are used is
governed by the rule for the particular n-phone context, as set up during rule
creation described in Section 4.2.2.
10The actual value of the point has to come from somewhere, but it is displayed in two profiles. For the
one that actually provides the value, the point is not marked as a phantom point. For the other profile,
the point is a phantom point.
9
•Posture composition-rules creation, management, and testing
–“Rule Manager”: specifying what combinations of postures are relevant when
choosing interpolation methods, as above. Rules may be created, edited, and
deleted, and used to specify which Transitions and Special Transitions should
be applied to which parameter in each segment of the n-phone represented.
–“Rule Tester”: in which symbol strings may be entered to check which rules
will apply. The rule displayed is the one consuming the most posture tokens
that fit a rule, starting at the left end. The postures may be shifted left, and
further postures added, allowing strings to be successively checked.
4.2.2 The Gnuspeech database that is edited using Monet
The database comprises:
1. posture names which may be assigned arbitrarily;
2. data associated with each posture to define the nominal targets, nominal durations
of components, the “beat” location11. All are accessible as meaningful symbols
chosen and specified by the language developer;
3. equations defining event times, accessible as symbols (which may be used in the com-
putation of additional symbols, or in the specification of points in the interpolation
templates);
4. parameter transition prototypes (interpolation specifications) which can be applied
in arbitrary ways to arbitrary individual parameters for arbitrary posture combina-
tions selected by combination rules designed by the user;
5. special parameter variation prototypes to manage added speech events that are
superimposed on the basic parameter tracks (for example, noise bursts, micro-
intonation, pitch excursions, and the like);
6. context sensitive rules to specify which prototypes apply to which parameters and
when;
7. meta-parameters (higher level parameters) that allow TRAcT parameter variations
to be derived from a higher-level representation framework, including such items as
tongue position, jaw rotation, lip rounding etc, based on further defined symbols,
derived symbols, and equations. The intention is to allow synthesis to be defined
on the basis of physical articulator movement which can fairly readily be converted
to the lower-level tube radii. Meta-parameters are not yet in use or even defined.
This would be a useful next step in developing the system.
These broad divisions correspond to the various subsystems and data structures that
together underlie Gnuspeech. They are created and modified by the Monet database
11The beat location is the time-location in a syllable where a subject will tap when asked to tap to
the rhythm of a sentence (Allen 1972a, 1972b). Allen found increased syllable stress increased reliability.
The beat times somewhat precede the onset of the nuclear vowel of the salient—stressed— syllables by
an amount correlated with the length of the initial consonant.
10
editing facilities. The overall database itself is keyed by the postures and posture combi-
nations that, accessed sequentially, create continuous speech. The databases in Gnuspeech
allow for the contextual effects of up to four adjacent postures (tetraphones). The system
could be modified to take more context into account, if necessary. For the current system,
this has so far proved unnecessary. Context-dependency is equivalent to using diphones,
triphones or tetraphones as a basis for synthesis, and allows various co-articulation effects
and specific consonant clusters to be handled effectively and efficiently. Context match-
ing is based on logical operations on combinations of individual postures (phones) and
categories of postures (such as nasal, or voiceless stop, in the current database).
5 Monet-based Speech Server access
5.1 Overview
Monet includes a facility for synthesising speech according to the current database, with
input either as plain text or as Monet-syntax symbol strings. The speech may be mono-
tone, or subject to intonation contours applied according to the modified Halliday (1970)
model—including or not including micro-intonation, or based on manually altered/created
contours. The speech output and Monet parameter tracks may be stored. The ability
to listen to speech during database development is an essential facility for testing the in-
tegrity of the data and rules being developed for a language, whether creating a Gnuspeech
language database from scratch, or simply modifying an existing database.
5.2 Segmental synthesis
Figure 3 shows the appearance of the Monet Synthesis Window during normal operation.
At the top are two text fields that allow plain text or posture stream syntax to be entered.
The “Parse Text” button on the right of the “Text” field causes plain text to be translated
and placed in the “Postures” field. Changing either field at this stage causes the other
field to go red, indicating the two fields are no longer equivalent.The check box on the left,
below the input fields, selects whether or not the parameters are stored. The remaining
buttons in the top area select “Synthesize” the text, and also allow the speech output to
be stored in a sound file.
The “Graphs” button on the right activates a pull-down panel that allows a choice to
be made of which tracks to plot. The default is for all tracks to be plotted and Figure 3
shows that these tracks are the dominant feature of the display. Scroll bars provide access
to different track areas, and allow the time record to be moved left and right as needed.
The filled circles represent the event times actually used by the rules to construct each
parameter track. Note that empty tracks may initially be displayed until the scroll bar is
used to bring tracks with plots into view. This “feature” should be improved!
Above the parameters track sub-panel are fields that display the “Time” and “Value”
of the cursor’s position in any of the parameter tracks. Below that two rows appear
above the parameter track displays showing the posture symbols for the nominal regions
governed by the n-phone posture rules, with the identification numbers of the rules used
in the second row regions. Below the parameter tracks are, from left to right: a button to
save the graph images; a button to show all the event values and times in a new window;
and a slider to change the horizontal scale of the displays.
11
Figure 3: The “Synthesis Window” showing 7 of the 32 parameters for the
first part of the utterance “Do you by chance happen to know what time it is?”
Thus the operation of the Synthesis Window is straightforward. The postures field can
be edited, and the changed string spoken, to see the effect of different posture selections
or changes rhythm and intonation.
Figure 4: The “Synthesis Window” showing a selected time extent and a time line
Apart from the use of the cursor to display the time and value at any point on a
parameter track, the synthesis window includes a line cursor (green) running over all
the tracks that can be moved to facilitate relating the points in one parameter track to
points in other tracks. By clicking the mouse within the tracks area, a line green cursor
appears, and can be dragged to any selected time position. Additionally, by shift-clicking
and dragging, a time range may be selected for synthesis. Figure 4 shows both the line
cursor, and a selected region. The two facilities are independent. Clicking on a time
position moves the line cursor to that position. Holding the mouse button down and
12
dragging allows the cursor to be dragged. Holding the shift key down whilst clicking
without moving allows the selected area to be cancelled or, if the mouse is dragged, the
selected area is changed to whatever time extent the new shift-click dragging covers.
5.3 Intonation and Rhythm
The rise and fall of pitch that together constitute “Intonation” is closely tied to the
rhythmic beat of an utterance. The salient syllables are what Halliday (1970) uses to
divide utterances into what he calls “feet” (see Appendix E), each foot beginning with a
salient syllable. The perceived rhythm of the speech is dependent on these beat times,
whose precise timing is, in turn, are determined by the durations of the sounds generated
from the posture sequences when speaking. The feet are grouped into what Halliday
calls “Tone Groups” of which he defines 5 major varieties, plus two combinations, and
a number of finer secondary variants (see Appendix E). The feet may be: “pretonic”,
leading up to the “tonic”; or tonic—said to be “marked” and are specially lengthened,
with the first foot of the tonic receiving the main intonational effects. Utterance final feet
are also marked. All postures in a marked foot have longer duration than in umarked feet.
Post-tonic feet, those following the tonic up to the beginning of the next tone group or to
the end of the utterance, are not marked except for the utterance final foot (which may
also be the tonic foot of the tone group. It was found that the rhythm and intonation
resulting from this simple model, created on the basis of extensive experiments in the
author’s lab, was perceived as natural as a more complex model due to Collier and ’t
Hart and their co-workers ’t (for example: Hart et al. 1990; Willems et al. 1988) in
subjective experiments (Taube-Schock 1993).
Our rhythm experiments showed that there were three independent sources of variation
in determining the duration of the component postures in feet. Thus these are the basic
determinants of rhythm—(1) the posture identity (approximately 45%); (2) whether the
posture was in an marked rhythmic unit—foot—or an unmarked unit (approximately
15%); and (3) a regression on posture length, decreasing—as the number of postures
in a foot increased (approximately 10%).12 Thus a total of roughly 60% of the rhythm
of spoken British English for synthesis is accounted for by choosing the durations of
the postures according to their marked or unmarked varieties13 with another 10% of
rhythm accounted for by the rhythmic unit regression. This allows the SpeechServer to
produce articulated speech with roughly 70% of the rhythm accounted for. This has a
very significant effect on the naturalness. The remaining 30% of the rhythm may well
represent free variation, as all of the traditional phonetic correlates of posture duration
variation were investigated and found not to be independent of the three factors above.
Other intonation experiments we carried out showed, amongst other things, that lis-
teners seem to perceive intonation changes “categorically”—in the linguistic sense of that
word. That is to say there were timing boundaries, and pairs of utterances with intona-
tion changes were only perceived as different if the intonation changes for the members
12This effect has misleadingly been called “isochrony”, implying equal length of the feet when the
length of different feet is, in fact, merely a little more equal than might be expected, based on the
nominal duration of the postures involved. Nevertheless, the effect is important (Hill, Witten & Jassem
1977, Jassem, Hill & Witten 1984), and is accounted for by a linear regression that reduces the length of
all the postures in a foot in proportion to the number of postures, as the number of postures in the foot
increases. The effect is insignificant for unmarked feet but very significant for marked feet.
13This leads to 132 postures comprising marked and unmarked versions of 66 different postures, in-
cluding silence, rest, and some artificial postures placed by the rewrite rules.
13
of the pair occured on opposite sides of that boundary (Hill & Reid 1977). Thus it is im-
portant to get the intonation changes in “the right place”, but provided such boundaries
are taken into account, the exact placement is not critical. However, utterances typically
have at least one tone group and its associated “tonic” foot. The postures in the tonic
foot have marked duration, and foot receives a major movement of the pitch, with the
tonic syllable—the initial syllable of the tonic foot—having the lion’s share. The tonic
feet provide the information points of utterances.
More work is required in these areas. We used a simplified version of Halliday’s
intonation model for Gnuspeech contours (implementing only statements, questions and
emphasis—factors that could be deduced from the punctuation), and implemented rhythm
based on: posture identity; whether the posture is marked or not; and how many pos-
tures occurred in a marked foot, as well as the location of the beat. Halliday’s complete
intonation model is outlined in Appendix E.
Currently, the tonic for the tone group always defaults to the last foot in the tone
group. Some additional parsing (or even better, understanding) could allow the tonic(s)
(information point(s)) of the utterances to be placed more intelligently, according to the
context for which the utterance is intended. For example, in the utterance comprising a
single statement tone group (tone group 1), in answer to a question about “the killing of
John”, the placement of the tonic or information point of the tone group (indicated by
bold italic below) depends on exactly what question is being asked:
1. “Bill killed John” answers the question “Who killed John?”
2. “Bill killed John” answers the question “What did Bill do to John?” and
3. “Bill killed John” answers the question “Whom did Bill kill?”
Figure 5: Comparison of less natural (Tone Group 1) and more natural (Tone Group13)
synthetic speech for the utterance “She was 98 when she died”
Listen to a comparison of the Tone Group 1 version followed by Tone Group 13.
14
5.3.1 The intonation parameters & the intonation window
Figure 6: The manual intonation
settings
Figure 6 shows “Intonation Settings” window
that produced the intonation contour portrayed
in the full “Intonation” window shown as Fig-
ure 7. The check boxes at the top show that
all the options were enabled. “Macro” refers
to the basic intonation contour; “Micro” en-
ables microintonation; “Smooth” applies a sim-
ple curve fit to the contour (which needs to be
improved); and “Drift” currently does little and
is a historical relic. “Tempo” allows the rate of
utterances to be change—1.00 represents nor-
mal speed with fractions being slower and >1.0
being faster. “Random Perturbations” applies
random changes to the value of points defining
the intonation contour to provide variety when
repeating the same utterance.
The intonation parameters used may be se-
lected as “Use Tone Group Models” or “Use
Manual Parameters” using the radio buttons.
In the latter case the manual parameters that
control the “Notional Pitch”, the range and
perturbations of the pretonic, and the range
and perturbations of tonic may be entered as
positive or negative floating point values. The
Notional Pitch determines the pitch at which
the intonation contour starts within the two octave range from -24 to +24, with zero
being Middle ‘C’ (261.626 Hz). The remaining four parameters use positive values for
a rise and negative values for a fall. When “Use Tone Group Models” is selected, the
manual parameters are greyed out and ineffective whilst “”Randomly Select Within Tone
Group” becomes active, and uses the basic tone group models, but allows both pretonic
and tonic point values and ranges to be varied somewhat from their nominal values in
successive utterances.
An utterance to be synthesised is entered into the text field of the Synthesis Window
and parsed (see Figure 3). The Synthesise button produces the spoken version. There is
also a Synthesise button in the Intonation window which also produces a spoken version.
However, the two buttons have somewhat different uses and effects and will be referred
to as “Synthesise-1” and “Synthesise-2”, or “S-1” and “S-2” for short.
S-1 always speaks based on generating an intonation contour for the utterance either
according to the settings from the models—perhaps with random perturbations, or ac-
cording to the manual settings, depending on which radio button was active. S-2 always
generates a spoken utterance using the contour that is on display, which can be one pre-
viously generated by S-1, or one produced by clicking the “Generate Contour” button on
the Intonation window; or it can be one that was modified from a previously generated
contour. Points defining the contour may be added, or deleted, or moved. To add a point,
Option-click at the time the point is required—the point will appear at the time and value
position of the cursor, dragging the contour towards itself as necessary.
15
Figure 7: The full intonation window
To move or delete a point, first “lassoo” it by dragging the cursor across it. Hitting
the delete key removes it. Using the arrow keys moves it in increments—a semitone at
a time in value—up or down, or by a distance related to the posture structure in time,
forward or backward. A red line appears if the forward or backward changes moves the
point on top of another point.
To achieve finer positioning in frequency of a selected point, the semitone value or Herz
value may be entered in the appropriate field under “Point values” at the bottom left.
To achieve finer positioning in time, a point can be added at whatever time is required,
positioning the cursor on the scale at the bottom when clicking, and then repositioning
the point to whatever frequency value is required.
The slope field is not currently very useful or informative and can be ignored for
now. The sub-panel on the right of the Point Values shows all the rules that were used
in the composition of the postures. The extents of the rules are marked on the plot of
the intonation as stripes alternating between white and grey. The dotted lines represent
potential beat locations for the utterance, which are identified along the top edge of the
plot. Above that, registered with the stripes, are the rules and the rule durations. A
scroll bar underneath the plot allows an oversized plot to be scrolled. If it is not visible,
just increase the height of the window a little and it will appear. There is also a “Scale”
slider just below the scroller, on the right. If the scale is reduced to accommodate a longer
contour, some of the rule durations may overlap and become unreadable. Enlarging the
scale and scrolling to the area concerned allows the information to be seen
At the bottom right corner are fields displaying the “Beat time”, the “Beat Offset” and
the “Absolute Time” of the point selected. If no point is selected, these are blank. They
are not editable (but they perhaps should be, to facilitate the absolute time positioning
16
of a selected point). Below those fields two buttons are placed: “Synthesize” (S-2) which
has already been discussed and “Generate Contour” which generates the same contour as
would be generated by clicking S-1.
When synthesising from the “Synthesis” window, the basic macro contour form is
generated (no smoothing) and the contour displayed varies in value, even when all six
check boxes at the top are unchecked. However, the speech output is monotone, as
required. It is debatable as to whether the contour should be a straight line, with the
points still shown. The contour that is produced varies appropriately according to which
of the two radio buttons is activated. If “Use Tone Group Models” is active, the “Manual
Parameters” lower down are inactive and greyed out. Also the “Randomly Select Within
Tone Group” is active and may be checked if desired. The effect of this is to change
the notional pitch, the pretonic slope, and the tonic slope within the basic tone group
framework. Selecting “Random Perturbations” adds random displacement of the contour
points for both the radio button selections. These displacements can be large enough
to change the perception of the tonic—audible competition if “Macro” has been checked
to produce non-monotone speech. This is not a good feature. The drift parameters are
greyed out if the ”Drift” box is unchecked, but Drift is not currently relevant.
If “Use Manual Parameters” is active, then the Manual Parameters are active and not
greyed, and the contour shows the effect even though the intonation of the speech output
is monotone. The “Randomly Select Within Tone Group” selection is greyed out and
inactive in this condition.
Note well, that the foregoing applies to synthesis using the S-1 button. If the S-2
button in the Intonation window is used, then the intonation applied always follows the
contour on display.
“Generate Contour” always generates a new contour matching the contour that would
be generated if the Synthesis window S-1 button were pushed.
It may be noticed that going for the Randomly Select Within Tone Group speech,
with Smooth checked, produces a fairly flat version of the contour, just pretonic and tonic
in straight lines. Adding smoothing actually does the opposite of smoothing. What it
really does is to introduce another source of variation, but it does avoid sharp corners
(first and second order discontinuities) in the contour. A better term for what “Smooth”
does might be summarised as “Wavy”. A more appropriate smoothing algorithm may be
appropriate, though it is hard to hear any difference between smoothed and unsmoothed
contours when Macro is enabled, except that if excursions become large enough they may
affect the perception of the contour.
17
6 Database creation and modification
6.1 Preamble
The Monet database editor is described in what follows as if it is fully working, which
it is not quite—yet. The port, well under way, needs to be completed. However, the
guide is not only intended as a manual for the use of the system, but also a guide for the
completion of the port and further development. The approach is intended to provide
insight into the real value of Monet as a linguistics tool—beyond the task of articulatory
synthesis. The Speech Server itself is fully functional, independent of the database editor,
and the parts of Monet that use the Speech Server for synthesis are complete.
6.2 Getting started
Figure 8: The Gnuspeech “Tools” menu
The major subsystems in Monet corresponding
to the above divisions/functions, and accessible
from the “Tools” menu in the bar at the top of
the screen and reproduced as Figure 8.
When developing a new database, or mod-
ifying an old one, the parameter computations
necessary for synthesis are carried out “on the
fly” (i.e. in real time) based on whatever ele-
ments of the databases are available at the time.
Thus the Monet system may be used to create
a new database file as well as to audit and edit
existing database files. The following descrip-
tions refer to the existing English database, as
defined in the file “Diphones.mxml” as noted
below in this section.
If you are using Gnuspeech under GNUStep
on a GNU/Linux system, you will see the
“Main” menu and can then select the “Panels” menu, as seen in Figure 9. However,
in that case, this manual is not the right manual for what you want to do, because the
interface for the GNUStep version of Monet is significantly different to the Macintosh
version. This manual is for the Macintosh version.
It is assumed that you have installed Monet on your system. If you intend working
on an existing database, you need to install the database file you wish to work on in
the Monet application “Resources” folder. If you intend working on the file supplied
with the application, you should also make a back-up copy in a safe place. So “Control-
click” on the Monet application and choose “Show Package Contents” from the pop-up
menu that appears to show the “Contents” folder. In that folder you will find several
items, one of which is “Resources”. Look in that folder and you should find a file named
“Diphones.mxml”. That is the file you should back up if you will be modifying it; or move
and then insert an empty template file from “Documents>Monet” if you wish to create a
completely new database. (The supply of templates for creating a new database has yet
to be created.)
The main dictionary, if you need to alter it, is located in:
Frameworks>GnuSpeech.framework>Resources>2.0eMainDictionary.dict
18
Figure 9: The “Main” and “Panels”
menus: GNU/Linux GNUStep
along with the list of suffixes for extending the
reach of the dictionary by allowing compounds
involving the suffixes to be found. For a lan-
guage other than English, these will need to be
replaced.
Most of the facilities you require to cre-
ate and modify databases are accessed via the
“Tools” menu selection (see Figure 8). How-
ever, to work on a language other than English
you will need to modify other components, in-
cluding the “Parser” in the Speech Server. This
module is a pre-processor that converts items
such as numbers, dates, special characters, possessive “’s”, and other such items into con-
ventionally pronounceable form. As this is built into the Monet code (it should be made
more accessible so that recompilation is not needed), you will need to change some of the
Monet code and recompile using the Macintosh xcode development tools if you wish to
change the Parser. Initially it can perhaps be ignored, but the task of creating a new
language needs to be made much easier.
6.3 Creating/editing the database
6.3.1 Categories, Parameters, and Basic Timing Framework
The Data Entry window allows the basic posture Categories, Parameters, Meta Param-
eters, and basic Formula Symbols to be defined. Figure 10 shows a partial window with
“Categories” selected, overlapped by a full window with “Parameters” selected. The
Categories selection allows categories into which postures might fall to be defined.The
categories can used, in parallel with individual posture names, when specifying the com-
position rules used to process posture strings in order to create the synthesiser parameters
tracks needed to speak the utterance. It is a way of generalising what would otherwise be
rather specific rules, and cuts down on the number of rules. It is coincidental that there
are 16 categories as well as 16 parameters. Categories may be added and deleted, and
comments may be supplied. The menu selection “Posture Categories” provides a window
showing to which categories each posture belongs (a posture may belong to several). The
comment at the bottom notes that the category “phone” cannot be deleted.
The Parameters selection allows the parameters required for the synthesiser to be
defined, as well as their minimum, maximum, and default values to be set. Comments
can be added for each parameter. Parameters may also be added or removed using the
buttons at the bottom right.
Figure 11 shows a portion of the same data entry window but this time with “Formula
Symbols” selected. The symbols shown and specified are used as the basic time framework
for compositing the posture parameters. The Minimum, Maximum, and Default timing
values are specified. Additional symbols may be defined in terms of these basic symbols,
plus the timing values derived from posture data, using the “Equations” selection in the
“Prototype Manager” (Figure 13 in section section 6.4). These are then used to define
detailed timing in the rule-based posture composition process.
19
Figure 10: Two Data Entry windows “Categories” & “Parameters” selected
Figure 11: Data Entry window “Formula Symbols” selected
6.3.2 Posture Editor
Figure 12 shows the “Postures” window that allows postures to be added, edited and
removed. To add a posture, click on “Add” then enter the name of the new posture and
hit Return. The posture will move to its correct position in alphabetical order. Find
it and click once on the entry (if you double-click you can change the name). With the
name highlighted you can use the selections “Categories”, “Parameters”, “Meta Param-
eters”, and “Symbols” to enter those data for the new posture. If starting from scratch,
the window will already have “phone” as a category (supplied by the template you ob-
tained), but other appropriate categories can be added by clicking the appropriate boxes.
The parameters will be set to default values but can be changed by selecting Parame-
ters, double-clicking the specific parameter value, entering the required value, and hitting
Return.14 Meta Parameters are not yet used but will operate in a similar fashion. The
Symbols selection allows different posture timing to be entered, based on changing the
default values that have already been entered, by double clicking the value, entering the
14The basic Parameter targets specify the nominal values needed to drive the synthesiser to realise each
posture. For our articulatory synthesiser database these include: cross-sectional radii, r1 through r8; the
state of the velum; specification of superimposed microintonation; and so on, as shown in Figure 10.
20
Figure 12: The Postures Parameter entry window
new value and hitting Return. Which of the symbols is used, and in what portion, is de-
termined by the composition rule that gets selected when a string is composited, but the
“duration”, and “transition” times are nominal values for the quasi-steady-state duration
and the transition to and from the posture quasi-steady-state, whilst “qssa” and “qssb”
represent the division of the quasi-steady-state into nominal first and second portions.
How they are actually used depends, as noted, on the applicable composition rules—and
their associated equations (see Section 6.4).
6.3.3 Prototypes, Equations, and Transitions
The data needed to make the entries described must be supplied from phonetic analy-
sis of natural speech in the target language and dialect, augmented by published data.
Separating the components for transition and quasi-steady-state portions requires judge-
ment and insight. The process is not strictly a traditional phonetic analysis which tends
to lump transitions with the overall duration of “vocoids” (open sounds with little con-
striction). The transitions between two postures will frequently be dominated by one of
the postures—for example, the stop-to-vowel transition in English depends mainly on the
stop rather than the vowel, though the details may depend on both. The composition
rules and transition prototypes (see Sections 6.5 and 6.4.2) have to be designed with these
constraints in mind, plus the fact that the transitions may start and end at different times
for different parameters.
The posture data and rules then require evaluation on the basis of speech synthesised
21
using the database being built. This process involves both spectrographic analysis of the
synthetic speech and careful listening trials which initially may be informal, but ultimately
require formal subjective-test evaluation.
In the current database for English, there are 132 posture variants (2 x 66 as previ-
ously noted). Some are closely related to conventional phonemes whilst some are special
postures, added by rewrite rules (see Appendix C.1), that are applied in the preprocess-
ing, along with the parsing of numbers and so on. The large number results from the
rhythm studies and the design of the intonation and rhythm models. (see section 5.3).
6.4 Prototype Manager and Transitions
Figure 13 shows the Prototype Manager. Three selections may be made: Equations,
Transitions, and Special Transitions
6.4.1 Equations
With “Equations” selected, as in Figure 13, named equations may be created to define
event times, in terms of the basic named time symbols of postures. They may be grouped.
New named groups may be created. Both equation names and group names should be
designed to indicate the nature of the association of the group members and the purpose
of event times corresponding to the named equations. The named event times are used
to determine where changes will occur when creating Transitions and Special Transitions.
The event times determined by the equations are created as needed. Superfluous equations
may be removed.
By defining some timing symbols in a regular succession, points in the Transition Pro-
files may be systematically moved by selecting different symbols for successive synthesis
trials of a given utterance. In this somewhat tedious manner, systematically varying stim-
uli may be produced for various purposes, such as psychophysical experiments (e.g. voice
onset time experiments). The system should be provided with more (semi-)automated
tools to facilitate the kinds of time and other changes needed for such work.
6.4.2 Transitions
Figure 14 shows the same prototype window, but with Transitions selected. The partic-
ular transition on display is “BasicContoidTriphone”, that is put in the Carr´e group of
transitions. The shape of the points on transition graphs indicate the segment to which
the transition leads. A “•” leads to the second target, a “▲” to the third, and a “∎”
to the fourth (as in a tetraphone). Note that this triphone actually ignores the second
target and thus only has “▲” symbols showing.
Slope ratios are generally used to control the transition shape by defining the relative
slopes in the (ordinary) “Transition Prototype” because of the way the parameter tran-
sition shapes are converted into actual parameter tracks. Unlike the Special Parameter
Transitions—where the requirement is simply to add an absolute value excursion onto
the particular parameters track—with the normal transitions it is necessary to apportion
the amounts of change that are to take place in the successive segments of the transition
without knowing (at the time the transition prototype is constructed) what the absolute
time or change in value will be. That is to say, the same prototype shape, which is the
important aspect of the normal transitions, has to apply to a variety of time and value
changes. This is also related to why the last point of an n-phone is a phantom point—you
22
Figure 13: Prototype Manager: Equations
can’t generate the complete actual transition until you have the next n-phone target to
give you the actual value that determines the absolute change that will be apportioned
according to the slope ratios for the transition.
As noted, slope ratios are not used for “Special Transition Prototypes” since they are
simply excursion superposed on the basic shape determined by the Transition Prototype.
Focussing on the shape of transitions is an effective method of modelling the constraints
of the articulators without having a physiological model with all the muscles, masses,
joints and so on.
The left-hand column shows whether there is a comment associated with a given
transition (none are shown in Figure 14 or related figures. The box on the right of the
names tells whether the transition is actually used in a composition rule or not. The
comments are displayed in the lower field on the right (accessing and displaying the
comments that exist does not currently work).
The transition shape and timing points are displayed in the upper field on the right,
time progressing on the x-axis and percentage of the total change demanded on the y-axis.
Note that and extra 20% of the range is allowed in case it is necessary to define an under-
or overshoot of the target’s nominal value.
In order to edit the transition, the user needs to click on the “Edit” button which
brings up the Transition Builder window shown in Figure 15. The time position of a
given point may be changed by selecting the point (drag across it to lasso it, when it
becomes surrounded by a lightly drawn square), and then selecting a different event time
from the list on the right-hand side. To change the shape of the transition, click on the
numbered area below the transition where the slope ratios are displayed. The selected
slope ratio become editable, with the range to which it applies obvious from the highlight
that replaces the grey field. Note that when a slope ratio is changed, all the segments that
are determined by slope ratios are changed, since the total change will now be apportioned
23
Figure 14: Prototype Manager: Transitions
differently when the actual parameter values are computed for the specific targets involved
in generating the parameter track for the parameter(s) to which the transition applies.
This is shown in Figure 16 which shows just the relevant part of the Transition Builder
window during and after a fairly extreme slope ratio change to the second segment of the
BasicContoidTriphone transition that was originally seen in Figure 15.
At present it is not possible to remove a transition that has been built even though a
button for that is provided. For this reason there are a number of “dud” profiles, usually
numbered to distinguish them, with “dud” indicating that they are not good transitions
so as to avoid confusion during database creation. Obviously this needs to be changed.
To create a new transition, the “Add” button in the Prototype Manager button should
be clicked. A new “Untitled” entry will appear in the list of transitions. Enter an
informative name in the style typified by existing entries (i.e. “camelCased”15). Hit
return, then click “Edit” whilst the new entry is highlighted. A Transition Builder window
will open showing a linear diphone transition will appear. The pull-down menu below the
new transition name at the top left can be used to change to a triphone or tetraphone.
Points may be added by double-clicking, and then choosing the correct type (“•”, “▲”,
or “∎”). The percentage position of the added points may be changed by lassooing the
target point and entering the value. The time position may be changed by selecting the
equation defining the desired new event time from the list on the right-hand side. This
where the process of creating transitions can get derailed and confused, as there is no
“Undo” button—obviously one should be added. At present, the bad transition must
simply be renamed “Dudn”.
15Camel casing is the practice of writing compound words or phrases such that each successive word
or abbreviation begins with a capital letter. Camel case may start with a capital or, especially in pro-
gramming languages, with a lowercase letter.
24
Figure 15: The Transition Builder window
Figure 16: Changing a transition shape
At the final boundary it will be nec-
essary to mark the point that defines the
end of the transition segment as a “Phan-
tom” point (as discussed above). There is
no apparent reason why this should not be
automated. As noted, the transitions pro-
vide appropriate change trajectories for the
different posture parameters in particular
phonetic contexts to replicate the move-
ments associated with articulation. It can
be helpful to have identical transitions with
different names associated with different
rules/contexts. For this reason a “Dupli-
cate” or “Alias” button would also be use-
ful to provide access to the same transition
via a different name (saving storage whilst
improving the human-computer interface).
The timing values that control the current transition segment are displayed at the
bottom left of the window. Since the shape of the transition is independent of the times
and values of any specific posture combination, the display is only for mis-information
(!) and is not currently activated—the same values are always displayed and changing
them has no effect. The actual timing is determined by the selected event times computed
from the timing associated with each posture and the current “Tempo”; and the absolute
values of the parameters depend on the posture targets perhaps modified in both timing
and value modified by the transition prototype.
25
In order to apply slope ratios to two or more transition segments in a new transition
it is necessary to select all the points that define the endpoints of the segments by Shift-
clicking them in turn, or lassoing them all at once. When all are selected the button
“Apply Slope Ratios” should be clicked. A new set of fields will appear beneath the
transition, initially with each slope ratio set to 1.0. Clicking on one of the fields will
highlight the selected ratio, with the visible field showing the extent of the segment that
is about to be changed, as describe at the beginning of this section. The port for this
facility has yet to be completed. Saving changes is not yet implemented.
Figure 17: Prototype Manager: Special Transitions
6.4.3 Prototype Manager: Special Transitions
The creating, editing and removing Special Transitions is very similar to the procedure
used for Transitions except that slope ratios are not applicable.
6.5 The Rule Manager
6.5.1 Introduction
Figure 18 shows the Rule Manager window with “Expressions” selected. These are often
called “The Rules”, and can be identified individually as “Rule n”. Each rule represents
a posture combination and will determine which transitions will apply to each parameter
transition for each sequence of postures covered by the rule, leading to the composition
of a further segment of the parameter tracks. There are currently 47 rules, including
“phone>>phone”, a rule which is always the last rule and cannot be moved or deleted.
Other rules can be both moved and deleted. The order is important in order to make
sure that the more specific rules are applied before more general rules.
26
Figure 18: The Rule Manager: Expressions
The rules are tried, starting from the top rule, and proceeding down towards the last
rule until a rule that “fits” a portion of the current posture string, starting with the last
posture used, is found. Thus the rules should be organised from the most specific at the
top to the most general at the bottom (phone>>phone)
Applying a rule to a set of postures may involve some rewrite rules.
6.5.2 Expressions
Selecting Expressions allows an expression comprising two, three or four postures to be
entered. Each posture may be specified explicitly by using a specific posture symbol, or
as a category, or as a logical combination of categories and postures. Consider Rule 7:
(silence) >>(b* or bx*) >>((((vocoid or nasal) or glide) or fricative) or silence)
Posture 1 is Silence. Posture 2 in the sequence can be a /b/ or /bx/16, and item 3
can be any single posture from the categories: vocoid, nasal, glide, fricative, or silence.
The second silence seems redundant because silence b* or bx*>>silence is not likely, but
complete coverage of possible inputs is important.
The individual postures sequences that are covered by the rule are listed in columns
below for convenience along with the number of legal combinations of postures. Note
that Figure 18 relates to rule 1, not to rule 7. Expressions may be added, modified or
16An “asterisk” (*) indicates a marked or unmarked posture—marked postures have a “prime” (").
27
removed using the appropriate buttons. To modify an expression, the components should
be entered in the Phone 1 to Phone 4 fields, and then the “Modify” button selected. At
present changes to the database cannot currently be saved because saving in the GUI-
based database editor is disabled. This is true of all the changes that could be made,
including those using the Rule Manager.
6.5.3 Symbols
When “Symbols” is selected, the named timing values determining the basic framework
according to the selected equations are displayed in the bottom left portion of the window
as shown in Figure 19 which presents a cropped portion of the Rule Manager window.
Figure 19: The Rule Manager: Symbols
A different equation may be used to determine these timing values of a given (selected)
rule by highlighting the value to be changed, finding the desired equation in the list on
the bottom right, and double clicking it. If a new equation is needed it is necessary to
go to the Prototype Manager (see Section 6.4) and create it. It will then show up in
the “Equation Names” in the Rule Manager window where it can be selected as just
described. Again, this aspect of the port remains to be completed. The nature of “Beat”
is explained in Section 4.2.2.
6.5.4 Parameters
Figure 20 shows the bottom portion of the Rule Manager when “Parameters” has been
selected. On the left are the names of the parameter Transitions that govern each of
the regular parameters for this rule. Highlighting a parameter and double clicking a
“Transition Name” on the right-hand side causes the selected transition to be used to
govern the transition(s) for that parameter. The updated Transition Name replaces the
one that was in place for the parameter previously. If a new transition is required, it
is necessary to create it first, using the Prototype Manager and the Transition Builder
28
Figure 20: The Rule Manager: Parameters
as discussed above (see Section 6.4.2).17 Note that each parameter can have a different
transition graph specified for each of the 47 rules.
The transitions are created according to linguistic knowledge, coupled with analysis of
relevant synthesiser output, produced in the various contexts covered by the rule that is
involved, using a high quality spectrographic analyser as well as informal listening trials
by native speakers. When some success in synthesis has been achieved, formal subjective
listening trials are called for.
The porting for this facility remains to be completed but much of the work is done.
6.5.5 Special Parameters
Figure 21 shows the bottom portion of the Rule Manager window with “Special Tran-
sitions” selected. The Special Parameters use the Special Transitions created using the
Prototype Manager and the Special Transition Builder.18 The procedure is very similar
to what was described for the Transition parameters management (Section 6.5.4). How-
ever, since these are special transitions, it is only necessary to enter the transitions that
actually occur. They are then applied by adding them to the basic parameter transition
during parameter track generation. In the example shown only excursions of “fricVol” and
“aspVol” are required to provide noise bursts for the n-phone rule involved. The Special
Transition can be set or changed by double clicking the appropriate name in the list of
Special Transition Names in the list at the bottom right. If a new Special Parameter
Transition is needed, it will be necessary to use the Prototype Manager and the Special
Transition Builder to create it (see 6.5.5). It will then become available in the list for the
Rule Manager: Special Parameters: Special Transition Name and can be selected.
17The Tools menu item shows “Transition Editor” instead of “Transition Builder”, but should not be
invoked by that path anyway. The same applies to the “Special Transition Editor” selection referencing
the “Special Transition Editor”. At present, the transition that comes up is an unalterable blank.
18Note the previously mentioned Tools menu name problem.
29
Figure 21: The Rule Manager: Special Parameters
6.5.6 Meta Parameters and Comments
The Meta Parameters selection currently does nothing, because no framework for handling
Meta Parameters has been formulated. The selection is provided for future implementa-
tion when such parameters have been formalised (for example, jaw rotation and tongue
height could be used to estimate selected tube radii; lip opening would affect R7 and R8,
probably as modified by jaw rotation).
The comments window applies to all the selections to allow suitable comments to be
added to document the reason(s) for a given choice, to note any special conditions, etc.
6.5.7 The Rule Tester
Figure 22: The Rule Tester
The “Rule Tester”, shown in Figure 22 al-
lows the user to enter a sequence of pos-
tures to see which rules will apply during
composition. When a sequence of postures
has been entered, clicking the “Test” but-
ton shows which rule would apply to the
postures and the number of postures used
for the rule, starting at the left-hand end.
The postures can be shifted left, keeping
the unconsumed tokens left over from the
previous rule, and further postures entered.
Currently the tester does not distin-
guish between marked and unmarked versions of the postures and will therefore not
operate correctly to distinguish two rules that differ only in whether marked or unmarked
postures are involved—if there were such rules. The notation in the rules showing an
asterisk (*) against a posture indicates that either a marked or an unmarked posture is
acceptable. A prime (') indicates a marked version
The basic timing parameters for the rule, Rule Duration, Beat location, and Marks 1,
2, and 3, are shown in the fields at the bottom right. Note that the time when changes in
30
the relevant transition shapes occur are derived from these, as defined for each transition
profile associated with each parameter.
7 References
ALLEN, GEORGE D 1972a. The location of rhythmic stress beats in English: an
experimental study I. Language & Speech 15(1), 72-100
ALLEN, GEORGE D (1972b). The location of rhythmic stress beats in English: an
experimental study II, Language & Speech 15(2), 179-195
ALLEN, JONATHAN, M SHARON HUNNICUTT & DENNIS KLATT (1987) From
text to speech: the MITalk system. Cambridge University Press: Cambridge, UK, ISBN
0 521 30641 8, 216pp
BOERSMA, PAUL 2001. Praat, a system for doing phonetics by computer.Glot
International 5:9/10, 341-347, accessed 2014-01-24)
CARRE, RENE & MOHAMAD MRAYATI (1994) Vowel transitions, vowel systems,
and the Distinctive Region Model. in Levels in Speech Communication: Relations and
Interactions. Elsevier: New York
CONDON, WILLIAM & WILLIAM OGSTON (1971). Speech and body motion syn-
chrony of the speaker-hearer. In D Horton and J Jenkins, eds. The Perception of Lan-
guage, pp. 15084. Columbus, OH: Charles E.Merrill.
COOPER, FRANK S (1957) Acoustic cues for the perception of initial /w, j, r, l/ in
English. Word 13 (1), 24-43
CENTRE FOR SPEECH TECHNOLOGY RESEARCH (undated). The Festival
speech synthesis system, http://www.cstr.ed.ac.uk/projects/festival, accessed 2015-07-11.
FANT, GUNNAR (1956) On the predictability of formant levels and spectrum en-
velopes from formant frequencies. In For Roman Jakobson. Mouton: The Hague, 109-120
FANT, GUNNAR & PAULI, S (1974) Spatial characteristics of vocal tract resonance
models. Proceedings of the Stockholm Speech Communication Seminar, KTH, Stockholm,
Sweden
GREEN, PETER S, (1959) Consonant-Vowel Transitions: a Spectrographic Study.
Studia Linguistica XII (2), Travaux de l’Institut de Phonetique de Lund, Bertil Malmberg:
University of Lund, 53 pages
HALLIDAY, MICHAEL AK (1967) Intonation and grammar in British English. (Janua
Linguarum, series practica, 48) Mouton: The Hague, 62 pages (David Crystal’s critical
review of the book is available as a 16 page download at:
http://www.davidcrystal.com/?fileid=-4919 (accessed 2015-08-11))
HALLIDAY, MICHAEL AK (1970) A course in spoken English: intonation. Oxford
University Press Australia/New Zealand, 134 pages (Out of print, but available in many
university and archival libraries.)
HALLIDAY, MICHAEL AK & WILLIAM GREAVES (2008) Intonation in the Gram-
mar of English. Equinox Publishing hard cover edition 2004, ISBN 978-1904768142, 237
pages
’t HART, JOHAN R, REN´
E COLLIER & ANTONIE COHEN (1990) A perceptual
study of intonation. An experimental-phonetic approach to speech melody. Cambridge:
Cambridge University Press
HILL, DAVID R (submitted) Low-level articulatory text-to-speech speech synthesis:
a working solution and a linguistics tool. (submitted for review to the (Canadian Journal
31
of Linguistics)
HILL, DAVID R, LEONARD MANZARA & CRAIG-RICHARD TAUBE-SCHOCK
(1995) Real-time articulatory speech-synthesis-by-rules. Proc. AVIOS ’95 14th Annual
International Voice Technologies Conference, San Jose, 12-14 September 1995, pages 27-
44
HILL, DAVID R, CRAIG-RICHARD TAUBE-SCHOCK & LEONARD MANZARA
(1992) Unrestricted text-to-speech revisited: rhythm and intonation. Proc. 2nd. Int.
Conf. on Spoken Language Processing, Banff, Alberta, Canada, October 12th.-16th.,
pages 1219-1222
HILL, DAVID R (1991) A conceptionary for speech and hearing in the context of ma-
chines and experimentation, http://pages.cpsc.ucalgary.ca/ hill/papers/conc/index.htm,
accessed 2015-05-06.
HILL, DAVID R, WIKTOR JASSEM, & IAN H WITTEN (1979) A statistical ap-
proach to the problem of isochrony in spoken British English. In Current Issues in
Linguistic Theory 9 (eds. H. & P. Hollien), 285-294, Amsterdam: John Benjamins B.V.
HILL, DAVID R, WIKTOR JASSEM, & IAN H WITTEN (1977) Some results from a
preliminary study of British English speech rhythm. Dept.of Computer Science Research
Report 78/26/5, U of Calgary,
http://pages.cpsc.ucalgary.ca/ hill/papers/british-english-speech-rhythm.pdf accessed 2015-
05-19.
HILL, DAVID R & NEAL REID (1977) An experiment on the perception of intona-
tional features. Int. J. Man-Machine Studies 9 (2), pages 337-347
HOLMES, JOHN N, IGNATIUS G MATTINGLEY, & JOHN N SHEARME (1964)
Speech synthesis by rule. Language & Speech 7, pages 127-143
JASSEM, W, DAVID R HILL, & IAN H WITTEN (1984) Isochrony in English speech:
its statistical validity and linguistic relevance. In Pattern, Process and Function in Dis-
course Phonology (ed. Davydd Gibbon), Berlin: de Gruyter, pages 203-225
KENYON, JOHN S & THOMAS A KNOTT (1944) A pronouncing dictionary of
American English. G & C Merriam Company, Springfield, Massachusetts, 484pp
LADEFOGED, PETER & DONALD E BROADBENT (1957) Information conveyed
by vowels. J. Acoust. Soc. Amer., 29 (1), January, pages 98-104
LAWRENCE, WALTER (1953) The synthesis of speech from signals which have a low
information rate. In Communication Theory, Butterworth: London, pages 460-469
LIBERMAN, ALVIN M, FRANCES INGEMANN, LEIGH LISKER, PIERRE DE-
LATTRE & FRANK S COOPER (1959) Minimal rules for synthesising speech. J. Acoust.
Soc. Amer. 31 (11), 1490-1499, Nov
van LIESHOUT, PASCAL 2003 PRAAT: short tutorial: a basic introduction.
http://web.stanford.edu/dept/linguistics/corpora/material/PRAAT workshop manual v421.pdf Univer-
sity of Toronto, Graduate Department of Speech-Language Pathology, Faculty of Medicine,
Oral Dynamics Lab(accessed 2015-07-11)
LISKER, LEIGH (1957) Minimal cues for separating /w, r, l, y/ in intervocalic posi-
tion. Word 13 (2), 256-267
MANZARA, LEONARD & DAVID R HILL (1992) DEGAS: A system for rule-based
diphone synthesis. Proc. 2nd. Int. Conf. on Spoken Language Processing, Banff, Alberta,
Canada, October, 12-16, pages 117-120
MATUDA, MARCELO (2015) Git repository access to Gnuspeech derivative Gnus-
peechSA, https://github.com/mymatuda/GnuspeechSA, accessed 2015-09-02
32
MATUDA, MARCELO (2015) Access to a Gnuspeech derivative simplified cross-
platform implementation of TRAcT, https://github.com/mymatuda/SimpleTRAcT, ac-
cessed 2015-09-02
MATUDA, MARCELO (2015) Simple interactive cross-platform Monet,
https://github.com/mymatuda/MonetCP, accessed 2015-09-02
MATUDA, MARCELO (2015) Modules to use GnuspeechSA with “Speech Dispatcher,
https://github.com/mymatuda/GnuspeechSASpeechDispatcherModule, accessed 2015-09-
02
O’CONNOR, JD, LOU J GERSTMAN, ALVIN M LIBERMAN, PIERRE C DeLAT-
TRE, & FRANK S COOPER(1957) Acoustic cues for the perception of initial
/w, j, r, l/ in English. Word 13 (1), 24-43
PULLUM, GEOFFREY, K & WILLIAM A. LADUSAW (1986) Phonetic symbol
guide. University of Chicago Press: Chicago & London, ISBN 0-226-68532, pbk, 264
pages
RICHARDSON, D, R DALE, & K SHOCKLEY; (2008) Synchrony and swing in con-
versation: coordination, temporal dynamics and communication. In: I Wachsmuth, M
Lenzen and G Knoblich (eds.) Embodied Communication. Oxford University Press
SMITH, JULIUS O (1987a) . Tech. Rep. STAN-M-39, CCRMA, Music Department,
Stanford University, CCRMA Technical Report STAN-M-39,
https://ccrma.stanford.edu/files/papers/stanm39.pdf, accessed 2015-07-24.
SMITH, JULIUS O (1987b) Waveguide filter tutorial. Proceedings of the 1987 In-
ternational Computer Music Conference, Champaign-Urbana, pp. 9-16, Computer Music
Association.
TAUBE-SCHOCK, CRAIG-RICHARD. (1993) Synthesizing Intonation for Computer
Speech Output. M.Sc. Thesis, Department of Computer Science, University of Calgary
TATHAM, MARCEL, KATHERINE MORTON & E LEWIS (2000) SPRUCE: Speech
Synthesis. in The Structure of multimodal dialogue II. eds MM Taylor, F N´eel & DG
Bouwhuis. John Benjamins Publishing Company, Philadelphia, Amsterdam, pages 271-
292
WILLEMS, NICO, REN´
E COLLIER & JOHAN ’t HART (1988) A synthesis scheme
for British English intonation. Journal of the Acoustical Society of America 84 (4) Octo-
ber, pages 1250-1261
YELAVICH, LUKE , JAN BUCHAL, TOMAS CERHA, HYNEK HANKE, MILAN
ZAMAZAL, C.M. BRANNON, WILLIAM HUBBS, ANDREI KHOLODNY (undated).
Speech Dispatcher, http://devel.freebsoft.org/speechd, accessed 2015-07-24.
33
A Some Monet posture parameter and phonetic data
Note: The Monet database contains the raw tube model parameters, such as tube radii.The following
information provides duration data for posture transitions and quasi-steady-states based on a conventional
phonetic analysis of Study Units 30 and 39—British English—from Halliday (1970). The parameter values
(formants and so on), which are not the tube model parameters themselves, represent a distillation from a
variety of sources, and were used to guide the derivation of the relevant tube model radii for the postures.
The noise frequencies and timing were used fairly directly. The latter formed the basis of the rhythm
model that was developed. See Hill et al. (1977). The tube radii for the Monet postures appear in
Appendix A.2 in the TRAcT manual.
Parameter
Phone Unmarked Marked qss values
tran qss dur tran qss dur vel F1 F2 F3 F4 FH2 BW Ax
m - 50 200 50 250 300 - (e) (e) (e) (e) - - -
h 30 49.7 - 50 66.5 - - (e) (e) (e) (e) - - -
gs 30 49.7 - 50 66.5 - - (e) (e) (e) (e) - - -
ot 20 6 - 20 10 - - 581 1381 2436 3500 - - -
aa - 58.8 84.1 - 96.8 122.1 - 748 1746 2460 3450 - - -
ah - 40.1 65.4 - 93.2 118.5 - 750 1500 2500 3500 - - -
a - 51.7 77 - 107.1 132.4 - 722 1236 2537 3500 - - -
e - 35.7 61 - 52.4 77.7 - 569 1965 2636 3500 - - -
i - 28 53.3 - 51 76.3 - 356 2098 2696 3700 - - -
o - 45.2 70.5 - 78.6 103.9 - 599 891 2605 3220 - - -
uh - 20.9 46.2 - 48.8 74.1 - 581 1381 2436 3500 - - -
u - 25.7 51 - 83.6 108.9 - 376 950 2440 3320 - - -
ar - 81.5 106.8 - 156.4 181.7 - 677 1083 2540 3410 - - -
aw - 88.8 114.1 - 168.8 194.1 - 449 737 2635 3700 - - -
ee - 57.1 82.4 - 116.5 141.8 - 285 2373 3088 3700 - - -
er - 106.9 132.2 - 142.4 167.7 - 581 1381 2436 3500 - - -
uu - 38.4 63.7 - 99.7 125 - 309 939 2320 3320 - - -
ah-uu - 20 112.4 - 40 148.2 - (f) (f) (f) (f)- - - -
e-i - 20 99 - 40 132.1 - (f) (f) (f) (f) - - -
o-i - 20 92.5 - 40 135 - (f) (f) (f) (f) - - -
uh-uu - 20 104.8 - 40 168 - (f) (f) (f) (f) - - -
in - 51.7 77 - 107.1 132.4 X 722 1236 2537 3500 - - -
an - 81.5 106.8 - 156.4 181.7 X 677 1083 2540 3410 - - -
on - 45.2 70.5 - 78.6 103.9 X 599 891 2605 3220 - - -
un - 20.9 46.2 - 48.8 74.1 X 581 1381 2436 3500 - - -
r 75.7 40.3 - 40 70.7 - - 240 1100 1300 2200 - - -
w 33.7 47.4 - 86.6 68.6 - - 240 500 2500 3320 - - -
l 22.6 72 - 58 84 - - 380 1500 3000 3700 - - -
ll 22.6 72 - 58 84 - - 309 1200 3000 3700 - - -
y 37.7 54.5 - 96.7 84.4 - - 285 2373 3088 3700 - - -
m 16 62 - 16 114 - X 190 950 2000 3320 - - -
n 25.7 60 - 25.7 88 - X 190 1850 3300 4200 - - -
ng 28.5 50 - 28.5 68 - X 190 2300 3400 4000 - - -
p 18.3 86 - 18.3 92 - - 190 460 2000 3320 - - -
t 24.2 60 - 24.2 80 - - 190 1780 3300 4200 - - -
k 30 82 - 30 104 - - 100 2600 2700 3500 - - -
ph 18.3 100 - 18.3 126 - - 190 460 2000 3320 - - -
th 24.2 100 - 24.2 118 - - 190 1780 3300 4200 - - -
kh 30 98 - 30 124 - - 190 2600 2700 3500 - - -
b 16 72 - 16 82 - - 100 460 2000 3320 - - -
d 18 58 - 18 86 - - 100 1850 3300 4200 - - -
g 14 60 - 14 80 - - 100 2300 2500 3500 - - -
bh 21.5 73.9 - 21.5 93.9 - - 100 460 2000 3320 - - -
dh 28.5 51.3 - 28.5 65.4 - - 100 1780 3300 4200 - - -
f 40 70.1 - 40 97.9 - - 100 690 2600 4000 1600 1000 0
th 40 54.6 - 40 116.1 - - 100 2000 2850 3950 1350 1000 0
s 29.6 78.1 - 29.6 111.5 - - 190 1300 3300 4000 6000 1000 0
sh 24.2 63.8 - 24.2 124.2 - - 190 2000 2700 4000 1500 500 0
v 44 48.4 - 44 68.9 - - 100 690 2600 4000 1600 1000 30
dh 44 80 - 44 108 - - 100 2000 2850 3950 1350 1000 30
z 30.4 56.5 - 30.4 84.8 - - 190 1300 3300 4000 6000 1000 30
sh 29.6 58 - 29.6 82.9 - - 190 2000 2700 4000 1500 500 30
ch 24.2 (118.5) - 24.2 (118.5) - - 190 2000 2700 4000 1500 500 0
j 24.2 (93.4) - 24.2 (100) - - 190 2000 2700 4000 1500 500 0
34
B Pronunciation Guide
B.1 Font comparison: Webster, Gnuspeech (Trillium), and IPA
Note: the posture symbols used to represent sounds throughout this document are usually “Trillium”
font symbols, except where they are enclosed in “/” pairs (for example /a "t a k/), representing a “broad”
transcription of the English word “attack”, using IPA symbols. Trillium symbols are the same as Gnus-
peech symbols, and follow IPA and Webster’s forms fairly closely, but they are closer to the normal
orthography of English. The following table provides equivalences, and some pronunciation help in the
context of Gnuspeech. The pronunciation examples use Educated Southern English (“RP”) pronuncia-
tion. The special case of the /r/ sound is discussed in Note 3 to the equivalence table that now follows.
Gnuspeech Specimen
Class Websters (Trillium) IPA IPA description Words
Short @uh @Schwa, (Lower mid above, banana,
vowels central, unstressed) collide, abut
"@ a2Lower mid central hud, humdrum,
(Pike) (See Note 1) abut
e e EUpper mid front head, bet, bed,
unrounded peck
i i ÌSemi-high front hid, tip, banish,
unrounded active
¨a oOLower mid back hod, nod,
rounded bottom, slot
˙u u ÑSemi-high back hood, pull,
unrounded full, look
Medium a aa æ Raised low front had, mat,
vowel unrounded apple, scrap
Long ¯e ee i High front heed, beat,
vowels unrounded machine, even
œ er @: Mid-central un- herd, birt,
rounded. Not US word, fertile
rhotacised version.
˙a ar ALow back un- hard, rather,
rounded. Not US artic, blah
rhotacised version.
˙o aw O: Lower mid-back all, gnaw,
rounded. Not US caught, worn
rhotacised version.
Diphthongs ¯a e i EÌ (See hate, today,
components grey, maiden
¯o uh uu 2Ñ (See hoed, boat,
components beau, crowed
a ˙u ah uu æÑ(See now, loud,
components bowed, out
˙oi o i oÌ(See boy, coin,
components ointment, noise
¯ı ah i æÌ(See nine, sight,
components buy, ply
35
Gnuspeech Specimen
Class Websters (Trillium) IPA IPA description Words
Glides & w w w Voiced rounded won, away,
liquids labio-velar waiver, always
approximant
y y j Voiced palatal year, yoyo,
central onion, aryan
approximant
r r r Commonly used zero, rise
for many “r”. Not arrow, ground
sounds. Not strict.
IPA. (see Note 2)
l l l Voiced alveolar let, alone,
lateral approx- lily, pull
imant (See note 3)
Unvoiced p p p Voiceless bilabial pat, slipper,
stops stop apt, piper
y t t Voiceless alveolar tap, wet,
or dental stop letter, potato
k k k Voiceless velar kill, lacky,
stop accent, cognac
Voiced b b b Voiced bilabial ebbed, banana,
stops stop rebel, pub
d d d Voiced alveolar udder, dad,
or dental stop elder, drop
g g g Voiced velar get, bigger,
stop hag, egregious
Nasals m m m Voiced bilabial me, mama,
stop lemon, dam
n n n Voiced alveolar now, canal,
or dental nasal enemy, train
Nng NVoiced velar ring, angst,
anger, unggulate
Unvoiced s s s Voiceless alveolar sit, sister,
fricatives central fricative whisper, juice
sh sh SVoiceless palato- shto, mission,
alveolar central quiche, action
laminal fricative fish
f f f Voiceless labio- fright, phone,
dental central effort, rough
fricative ruff
th th TVoiceless inter- thin, anther,
dental central truth
fricative geothermal
36
Gnuspeech Specimen
Class Websters (Trillium) IPA IPA description Words
Unvoiced s s s Voiceless alveolar sit, sister,
fricatives central fricative whisper, juice
sh sh SVoiceless palato- shto, mission,
alveolar central quiche, action
laminal fricative fish
f f f Voiceless labio- fright, phone,
dental central effort, rough
fricative ruff
th th TVoiceless inter- thin, anther,
dental central truth
fricative geothermal
Voiced z z z Voiced alveolar zip, zoom,
fricatives central fricative azalea, rose
zh zh ZVoiced palato- measure, Asia,
alveolar central corsage, beige
laminal fricative fish
v v v Voiced labio- vat, verb, over,
dental fricative avenge, rave
th dh DVoiced inter- that, mother,
dental central clothed
fricative then
Unvoiced ch ch ÙVoiceless palato- chat, fetch,
affricate alveolar affricate ratchet, church
Voiced j j ÃVoiced palato- jot, page,
affricate alveolar affricate judge, adjacent
Aspirate h h h Voiceless glottal- hat, house,
central fricative behind, haha
(See note 4)
Note 1: Same as schwa in many US dialects.
Note 2: English has “clear l” and “dark l” (or “velarised l”). Synthesisers may account for the
difference by a rewrite rule or composition rule since in English these sounds do not distinguish
words. The “dark l” occurs in post-vocalic positions (loosely, following vowels, diphthongs and
triphthongs). “Clear l” occurs elsewhere.
Note 3: In British “RP” English the “r” sound is pronounced much the same as in General
American in words like “zero” and “ground”. However, in a word like “flower” the “r” in not
pronounced as an “r” but a schwa vowel is used instead unless the word is followed by another
beginning with a vowel. In the word “herd” the “r” is omitted altogether. These are a typical
feature of “r” in British accents generally.
Note 4: In English, “h” is usually at least partially voiced in intervocalic position. Although
there is a distinct IPA symbol for this (“h” with a right hook on the tail), the effect may be
taken care of by a rewrite rule, as is appropriate for the “clear l”/”dark l” distinction of note 2.
Such allophonic distinctions should be taken care of by rewrite rules and the composition rules.
37
Note 5: In the Gnuspeech standard dictionary and spoken output, the pronunciation of all
words assumes a rhotic accent: that is, an “r” appearing in the orthographic form before a con-
sonant, or a place where a pause will occur when spoken, is pronounced, as in General American,
and unlike the Educated Southern English (RP) accent from Britain. Another systematic char-
acteristic of General American compared to the RP accent is the use of short /æ/ rather than
long /A/ in words like “command” and “dance”. In fact the Educated Southern English accent
seems to be changing in that direction anyway. Otherwise the Gnuspeech dictionary broadly
follows the RP accent as specified by the new Oxford English Dictionary, and as informed by
native speakers with British “RP” accents (typically heard on al Jazeera these days!). It is
considered that this gives an acceptable, if slightly strange mid-Atlantic accent. Later versions
should allow selection between more precisely defined, better approximated accents by switching
dictionaries.
B.2 Illustration of some pitfalls between American and British
English
Note: The underscore character (“ ”) is used to separate individual posture (phonetic)
elements and (later in Appendix B.3) a period separates syllables.
GA RP
Websters Gnuspeech IPA Example words Websters Gnuspeech IPA
h w h w h w when, whisper w w w
@r uh r @r herd, bird œ er @:
¯e r ee r i r cheer, hear ¯e @ee uh i @
e r e r Er error**, merry e r e r Er
e@r e uh r E @ r care, bear e @e uh E @
¯e @r ee uh r i @r leery, eerie ¯e @r ee uh r i @r
˙u r u r Ñr poor, moor ˙o aw O:
˙u r u r Ñr Lourdes, tour ˙u @u uh Ñ @
˙o r aw r O: r cord, lor ˙o aw O:
¨a r o r Or porridge, foreign ¨a r o r Or
** The second syllable in “error” is schwa r in GA and schwa by itself in RP.
This small collection of words indicates some of the different ways /r/ is treated in the two
dialects of English—General American (“GA”) and Educated Southern English (“RP”). How-
ever, a proper treatment of the vagaries of /r/, which includes “/r/ insertion” or the “intrusive
/r/” is beyond the scope of this manual. Suffice it to say that when a word ends in a vowel,
including the RP schwa vowel that usually replaces /r/, and the next word begins with a vowel,
an /r/ is inserted (for example, “The very idea of it!” becomes “The very idea-r-of it!” when
spoken by someone with an /r/ inserting dialect such as RP. This then leads to what is called
“hyper corrective intrusive /r/”, and condition in which an/r/ is inserted regardless. This is a
largely American peculiarity whereby someone with a traditionally non-rhotic accent (as found
in New York City and New England) hypercorrects and pronounces /r/ regardless of whether or
not it precedes a vowel. Hence we get “I’ve got no idear what to wear!” and “He liked to drawr
cats.” (http://dialectblog.com/2011/09/10/intrusive-r/ accessed 2015-05-30). Then there are
accents such as the Scottish accent for English that produces an alveolar trilled /r/ in words
like “porridge” (a very Scottish word, which uses the real IPA /r/), and the French uvular “r”
variant whose symbol is a turned (that is upside down) “R” /K/, not to mention the single-
38
flapped Spanish “r” as in “pero” that is represented by what is called a “fish-hook-r” /R/.
Like quantum theory, if you think you understand the “r” sound, you don’t!
B.3 Syllables, stress, and American versus British English
In what follows, underscore is used to separate Gnuspeech posture symbols with a “period”
(full-stop) instead to indicate nominal syllable boundaries. The standard Websters and
IPA symbols are used where appropriate. A stress mark (") denotes that the following
syllable is given primary stress. A secondary stress mark () indicates secondary stress.
Monosyllabic words in English are generally given primary stress only if they are “content”
words (that is to say, they are a noun,verb,adjective or adverb). Form words (the rest)
with only one syllable are not given stress (though some particular utterances may demand
contrastive stress such as: “We were on our way to the stadium, not from it”).
American English speakers do not all speak the same way, nor do all British English
speakers. Even within a group of people who nominally have the same accent, there
will often be individual variation. The topic of accent and dialect cannot be covered
here, except to draw the reader’s attention to the fact that the precise choice of the
sounds and stresses are among the factors that comprise an individual’s accent. Below
are represented a few words in a typical British and a typical American accent. These
examples hardly begins to address the topics of rhythm and intonation which are also
important, and (for English) closely tied in to stress and vowel quality choices. It also
ignores the more subtle differences in vowel quality between vowels which are represented
by the same broad transcription symbol, but are articulated somewhat differently between
(say) General American and RP. Narrow transcription (shown by enclosing the symbols
within square brackets—“[” and “]”—and adding “diacritical” marks) together with pho-
netic/phonological training—are necessary for real precision. Exactly how to represent
and reproduce correct vowel quality can still cause debate and confusion! Here are a few
words that are differently pronounced in GA and RP.
Word Gnuspeech
Accent Websters (Trillium) IPA
Polygon
RP "p¨a.li.g@n"p o.l i.g uh n "pO.lÌ.g@n
GA "p¨a.li."g¨an"p o.l i.g o n "pO.lÌ.gOn
Polygonal
RP p @."li.g@n@l p uh."l i.g uh n uh l p@."lÌ.g@n@l
GA p @."li.g@n@l p uh."l i.g uh n uh l p@."lÌ.g@n@l
About
RP @."ba ˙ut uh."b ahuu t @."bæÑt
GA @."ba ˙ut uh."b ahuu t @."bæÑt
Fire
RP "f¯ı@ "f ahi uh "fæÌ@
GA "f¯ı@r"f ahi uh r "fæÌ@r
Command
RP k@"m˙and k uh."m ar n d k@"mAnd
GA k¨a"mand k o."m aa n d kO"mænd
39
C Rewrite rules for Monet
C.1 Idiosyncratic rewrite rules and allophonic requirements
Rewrite rules are applied before the process of compositing the posture strings into pa-
rameter tracks takes place. A good example of the kind of reason they are applied is the
re-write that transforms an “l” followed by a contoid into the dark “l” allophonic variant,
or the rule that puts a glottal stop in-between the vowel at the end of one word if the
next word begins with the same vowel. More rules are probably needed.
[stop] >>[h* Šhv*] becomes [stop] >>q?* >>[h* or hv*]
[stop] >>[stop] >>[stop] becomes [stop] >>[stop] >>q?* >>[stop]
[affricate] >>[stop Šaffricate Šhlike] becomes [affricate] >>qc* >>[stop Šaffricate —
hlike]
[l* & end-of-word] >>[contoid] becomes [ll*] >>[contoid]
(We may need a similar rule for r to rr, but at present r and rr are the same)
[affricate] >>[stop] >>[stop Šaffricate Šhlike]
becomes [affricate] >>qc* >>[stop] >>q?* >>[stop Šaffricate Šhlike]
[vowel(i) & end-of-word] >>[vowel(i)] becomes [vowel(i)] >>gs* >>[vowel(i)]
(That is, the glottal stop only gets inserted for same vowels in succession)
We need additional re-write rules to deal with “the” before a vowel (it becomes “thee”)
and dealing properly with RP use of linking /r/ and intrusive /r/, basically avoiding the
mid-Atlantic accent with the inappropriate rhotic ˚ and sticking to what is called in
America an “English or Eastern accent (see the next section).
C.2 Linking /r/, intrusive /r/ and glottal stop insertion
From ⇓to ⇒aa ah a e i o uh u ar aw ee er uu ah i ah uu e i o i uh uu
aa 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
ah 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
a1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
e1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
i- - - - - - - - - - - - - - - - - -
o1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
uh 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
u1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
ar 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
aw 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
ee - - - - - - - - - - 1 - - - - - - -
er 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
uu - - - - - - - - - - - - 1 - - - - -
ah i - - - - - - - - - - - - - 1 - - - -
ah uu - - - - - - - - - - - - - - 1 - - -
a i - - - - - - - - - - - - - - - 1 - -
o i - - - - - - - - - - - - - - - - 1 -
uh uu - - - - - - - - - - - - - - - - - 1
Note 1:Kenyon & Knott (1944, page xxxv) also tell us that Eastern & Southern
American drops utterance final “r”after the same vowels. When visiting Los Angeles as
a late teenager, I delightedly remember being asked for the time by a local. When I
replied (“It’s five and twenty past five.”) he said you must be Canadian (we were actually
British). “Yes,” I replied, disingenuously, but how did you know.” “Well,” he said, “you
have what we call an English or Eastern accent.”
40
Brad Hodges phoned me about the tape I sent him. He was impressed, but he said
our use of “r” sounds like a Chicago teenager before maturity! That is our choice of a
“mid-Atlantic” accent for Gnuspeech! Bad choice, in my opinion.
Note 2: We could probably dispense with the diphthongs if the second components
were easily available for rewrite purposes.
D GnuspeechSA Speech Server command line tool
D.1 Introduction
Gnuspeech speech synthesis is available from the command line, or from applications
using GnuspeechSA (GnuspeechStandAlone). GnuspeechSA is a port to C++/C of
the TTSServer from the original Gnuspeech (http://www.gnu.org/software/gnuspeech/)
source code written for NeXTSTEP by Marcelo Matuda. It is a command-line program
that converts text to speech and may be obtained at:
https://github.com/mymatuda/GnuspeechSA
as well as this site. It doesn’t have external dependencies, apart from the system C++/C
libraries. In GNU/Linux applications users may use GnuspeechSA via Speech-dispatcher
(Yelavich et al. undated). Marcelo has also produced other free software versions of some
Gnuspeech components, including a simplified TRAcT, that run on Microsoft “Windows”,
available at:
https://github.com/mymatuda/GnuspeechOnWindows
D.2 Installing GnuspeechSA on the Macintosh
Compilation of GnuspeechSA on the Macintosh, requires xcode 4.3.2 or higher installed
to be able to use the Clang C++11 compiler and libc++ library module. Xcode is
a free download from the App Store. It is also necessary to install CMake 3.3.0-rc2
available at http://www.cmake.org (accessed 2015-08-22), which has a nice GUI. Finally
you need to clone the Git repository copy of GnuspeechSA onto your system and make
a “build” directory in the top level directory. Figure 23 shows the CMake window with
the source and destination fields entered, after clicking “Configure” and then“Generate”
when running under OS X 10.4 (“Yosemite”).
The warning message concerning the MACOS PATH that appears (in red) can be
ignored. A gnuspeech sa.xcode file will have been generated in the build directory. If not
using using OS X 10.4 (“Yosemite”) some parameter adjustment may be needed.
Double click the gnuspeech sa.xcode file to open Xcode and compile the sources. When
the compilation is complete a “Debug” directory will be found in the build directory that
contains a number of files, including the gnuspeech sa executable. If you change to the
build directory and type the following command19:
./gnuspeech\_sa -v -c ../../data/en -p trm_param_file.txt
-o do-you-happen.wav "Do you happen to know by chance what time it is?"
19Note: these two lines that follow actually represent a single line with a space at the line-break
instead of the new line.
41
Figure 23: The “CMake Window”—source & build directories added
followed by a Return. A .wav file of the spoken version of “Do you happen to know
by chance what time it is?” will appear in the Debug folder after significant diagnostic
output on the Terminal. To place the executable in its own directory, simply move or copy
gnuspeech sa, along with the “data” directory and the trm param file.txt. Placement of
the output file is determined by the “-o” parameter on the command line.
Here is an audio recording of the output generated by the command line above, when
executed in the Debug directory.
E Outline of Halliday’s British English intonation
notation
(The following summary is based on Halliday (1970). The book is out of print, but used copies
are available in many libraries and a few for sale—at Amazon, for example. Other relevant
references to Halliday’s work appear in the “References” section (Section 7).
E.1 British English Intonation
Natural conversation in English involves continuous selection from five tones. Subdivisions
depend on the “delicacy” used/required, which depends in turn on the requirements of
the grammar. However, Halliday does not tie phonology directly to grammar—English
is unfortunately (or more likely fortunately) not that simple. At the same time it is not
enough to treat intonation systems as the emotional icing on the communication cake.
42
Intonation contrasts are not lexical. Intonational and non-intonational aspects of
grammar co-exist (Halliday 1967; 2004). Given that features of an intonation system may
satisfy a variety of systems in the grammar, we should not fall into the trap of setting up a
different intonation system for each role. This would add needless confusion. The concept
of “tone” is an abstraction that provides a meaningful framework for the variations that
occur in natural speech.
For purposes of characterising intonation, four types of unit are recognised:
•the tone group, whose boundary is also a foot, syllable and phoneme boundary;
•the foot—the basic rhythmic unit;
•the syllable; and
•the phoneme—closely related to the class of sounds produced by a particular pos-
ture.
Following Abercrombie and Halliday, each foot begins with an accented syllable—the
beat (see Allen 1972a; 1972b) and may be followed by weaker syllables. Halliday claims
that some feet begin with a “silent stress” but, if so, this missing beat was seen to have
zero duration in our analyses. Successive feet make up tone groups—of which there are
five main categories, as noted, and in each simple tone group there is one obligatory foot
for which the initial syllable—the “tonic syllable”—receives “tonic prominence”, making
that foot the “tonic foot”. The tonic represents the “information” point of the phrase,
sentence or utterance. Feet occurring prior to the tonic, if any, make up the “pretonic”
(which is optional), and feet after the tonic, if any, simply follow the tonic pitch movement
at a reduced rate. The foot structure is determined by the occurrence of stressed syllables.
In Gnuspeech, the dictionary includes an indication of which syllable in a given word can
carry such stress (called “word stress”). Generally words that can carry word stress
are “content words”—nounds, adjectives, . . . , as opposed to the grammatical glue words
known as “form words” (words such as “and”, “the”, “of”, and “to”). Syllables in form
words and other unstressed syllables may be stressed for particular purposes (e.g. “But
what is the reason you went.”, “I said report, not deport.” This technique is overused in
North American sports reporting—the terms “offense” and “defense” are good examples.
The tonic may also be misplaced as when the ferry staff announce that “This is an
important safety announcement.” instead of “This is an important safety announcement.”
The tone groups are associated with the grammatical structure. They help to enhance
the intelligibility of the utterances, partly through their rhythmic effect, and they also
significantly affect their meaning at several levels—semantic and pragmatic.
The scheme for rhythm has usually been labelled as a “theory of isochrony”, probably
on the pre-instrumental perception that the beats occurred at more or less regular inter-
vals. Objectively, the beats do not occur at regular intervals. In fact the intervals may
vary in length by a factor of 7:1 between the shortest and the longest. However, our work
has shown clearly that—at least for the British English RP accent—the tonic and final feet
do have durations in which the feet with more phones are somewhat shorter than would
be expected from the natural (statistical) durations of the constituent phones (see Sub-
section 5.3). Thus there is some justification for calling English—particularly British RP
English—a “stress-timed” language in contrast to French, which is conventionally called
a “syllable-timed” language, where syllables tend to be perceived as occurring at regular
43
intervals. Listening to native speakers of these two language bears out this distinction,
at least at the perceptual level.
In the tonic foot there is always pitch movement, and the greatest rate of change of
pitch occurs on the tonic syllable. Halliday says that pretonic pitch is be “level” in some
pretonics, but in practice there is some pitch movement all the time in all speech, so level
is a relative term. The tonic and the pretonic are the only two places where a contrast
between tone groups can be made.
Figure 24 shows the five basic tone groups and their variants diagrammatically. There
are also two compound tone groups—13 and 53—that are made up of two successive
simple tone groups with no intervening pretonic; specifically tone group 1 followed by
a tone group 3; or tone group 5 followed by tone group 3. The second tone group in
either case is the minor one. The major information is carried by the first tone group—
tone group 1. The information in the second tone group (3 or 5) is subordinate to the
information in the first, though related.
There are three distinct areas of choice in producing the intonation for an English
utterance.
1. “TONALITY”: the distribution into tone groups—number and location of the bound-
aries;
2. “TONICITY”: the choice of tonic syllables; and
3. “TONE”: the choice of primary and secondary tone patterns.
Natural tonality defaults to roughly one tone group per grammatical clause—the
boundaries may not exactly coincide—and represents the distribution of information units.
Tonicity is constrained by the rhythm—the position of salient (stressed) syllables,
and the speakers intention to convey the information point(s) of what is being uttered.
Tonicity is clearly related to tonality. The neutral form places makes the first syllable of
the last element of the grammatical structure, which corresponds to the first syllable of
the last foot, the tonic syllable. A regular departure, deserving special mention is in “wh-”
or information questions. The tonic them falls on the stressed syllable of the interrogative
element(s).
An explanation of Tone, the different tones as illustrated in Figure 24, is necessar-
ily more complex and can only be partial, asking “what grammatical systems do these
tones expound”, and presupposing an explicit grammar. The basic frame breaks tone
into three moods, declarative, interrogative, and imperative, contrasted with “moodless”
minor clauses [“neutral” as opposed to marked??]. Tone 1 may be regarded as the neutral
tone, for declarative statements, with variants for simple statements, listing statements,
emphasis, and so on. Interrogatives are broken into polar—yes/no—and non-polar—
“information” questions. Yes/no questions involve a rising pitch in the tonic, but the
tones involved are subject to subtle choice depending on the situation and informational
context. Halliday gives a detailed exposition of when to use the various tone groups on
the basis of categorising clauses/utterances into: statements; questions of the two kinds;
commands; answers; and exclamations.
44
Figure 24: A summary of the five tone groups and variants comprising the Halliday
(1970) intonation system
45
The reader is referred to Halliday’s (1967) monograph for the full treatment, but this
summary indicates the flavour. Even the full treatment is only a guide to the bones of
British English intonation. Halliday says:
Tone marks a kind of activity involved, by a complex pattern built out of simple opposi-
tion between certain and uncertain polarity. If the polarity is certain, the pitch of the tonic
falls; if uncertain, it rises. Thus Tone 1 is an assertion, or a query not involving polarity;
and Tone 4, which falls and then rises, is an assertion which involves or entails some query.
Tone 2 is a query, 2 being a query about a specific assertion; and tone 5, which rises and
then falls, is a dismissed query, one countered by an assertion. Tone 2 avoids a decision; as
an assertion it is at best confirmatory, contingent or immaterial.
In his review of Halliday, Crystal criticises Halliday’s thesis as based only on assertion,
suggesting that Halliday has not shown that intonation functions systemically in the same
sense as grammar—see the Halliday (1967) citation. However, as a native speaker of
RP English who has studied Halliday’s work, and has been involved in the algorithmic
synthesis and recognition of computer speech, it is clear to me that intonation and rhythm
are essential components of meaning, and changes in intonation and rhythm dramatically
affect both the basic meaning of utterances and the nuances that a speaker wishes to
convey, not only by choice of words, but by choice of tonicity,tonality and tone, as well as
rhythm. The problem remains concerning how to systematise the use of these feature of
English. Halliday has made a giant step in that direction, but for high-quality synthesis
and recognition by machine, algorithms to untangle and use both meaning and speaker’s
intent are crucial. Punctuation alone doesn’t cut it.
46

Navigation menu