This book is about spoken language output -- how to build a synthetic
voice that can express a range of concepts in a given domain, whether
a system has a small range, and so an easily describable domain, or
an unlimited set of possible inputs.
The Festival Speech Synthesis System, and more specifically, the
FestVox voice building
tools, provide the framework for the discussion. All these
tools, Festival, and all other applications and databases given
in examples here are available as open source and/or free software.
This version is still a very early draft, it rondom changes from high
level discussion to low level code aspects without the appropriatae
introduction of the material. As it stands this version is most
useful to people with some speech processing experience who want to
build voices in the Festival Speech Synthesis System, and have some
patience to work through these notes. These sections
will become more general introductions in later versions.
Over the last several years, speech technology, as well as the
computational resources needed to carry out the necessary (or popular)
algrithms, have advanced considerably. We can build systems that
interact through speech; they can listen to what was said using speech
recognition, compute or do something, and then speak back, using
spoken language generation.
Language is a fundamental part of everyday life. Whether we are using
speech, sign language, or perhaps a coding system that conveys
meaning through touch, we use language to express our thoughts,
intentions, reactions, and experiences -- often scarcely considering
that we are even speaking, but rather feeling as if we were engaging
in direct manipulation of the notions with our conversation partner.
The complexity of the underlying systems involved in speech rarely
enters our minds as we go about our business, and it seems as if this
fosters an unspoken attitude: "It's easy to talk, it should be easy to
make a machine that talks." By the time you finish this book, you
should be aware of the components of the process, the relative effort
and expertise involved in each component, and where the trade-offs lie
in speech synthesis today. Our hope is that you will take away, even
with a light reading, an appreciation of the fundamental aspects of
speech communication, and how one might, with this modality, make
machines more useful.
speech.
The book is organized into six major parts: Overview and Use of
Speech Output, Building Synthetic Voices, Interfacing and Integrating,
detailed Recipes for building voices, and our concluding remarks
critiquing the state of the art, and discussing a number of interesting
issues that remain open in speech science and synthesis technology.
Part I
is a speech science and technology overview, primarily as it relates
to speech-interactive systems, and how to use speech synthesis. We
will be using the Festival Speech Synthesis System throughout the
book, and it is presented here. If you are familiar with speech
synthesis, you might want to skip to the discussion of Festival in
Chapter 3; if you're familiar with Festival,
you may wish to go directly to Part II.
In Chapter 1, we start with an introduction to
speech in general, the role of spoken language generation, and in
particular, of the basic issues in speech synthesis: text analysis,
prosody, and audio waveform generation; then in Chapter 2, we discuss speech science and technology in
greater detail, introducing the basic tools, concepts, and terms
involved. Chapter 3 is on using the Festival
Speech Synthesis System -- as a system, its basic use, and the modules
and architecture. The utterance structure we use is also discussed
here.
Part II
is on Building Synthetic Voices; here, start looking at the process of
building a voice, and illustrate it with demonstaions and
walk-throughs using the FestVox voice building tools. Chapter 5 covers limited domain synthesis, in which high
quality can be achieved at the expense of flexibility, if you know
what the system is going to say in advance. In Chapter 6, we then go into general synthesis, which
includes Text-to-Speech (TTS); the system should be able to speak, no
matter what it is given, with some expected level of quality -- even
if given only a string of text. It is in this chapter that we expose
most of the components of a general purpose synthesizer, such as the
text processing module, prosodic models that predict tune and timing,
lexicons and pronunciation modeling, and issues with data collection
and recording in order to create voices.
Chapter 10 is on various techniques for audio
waveform generation, once the system knows the final parameters of the
utterance. Evaluation and tuning of voices is discussed in
Chapter 14.
In Part III, we get to Interfacing and Integration issues.
Chapter 15
covers mark-up languages, or mark-up components of
standards, such as SABLE, JSML, and VoiceXML, and how these can be
used and implemented for synthesis. Then, in
Chapter 16
we go into some detail on Concept-to-Speech systems, and so touch on
language generation and the benefits of coupling the language
generation and synthesis components in terms of spoken output quality.
Chapter 17
is a discussion of deployment issues, such as client/server
issues, the memory and disk footprint, runtime constraints, scaling up
and pruning systems down to meet the available resources, and signal
compression.
Although step by step examples will be discussed in line throughout
the various chapters, Part IV offers
complete detailed walkthoughs of the actual steps involved
in a number of larger examples.
Chapter 18 gives a complete example of building a
Japanese diphone synthesizer, and Chapter 19
covers UK and US English diphone synthesizer while
Chapter 20 covers a the deisgn and executation
of building a limited domain synthesizer for ...
We follow with Part V, with a discussion of
the state of the art and the future of spoken language output, and then
a set of appendices.
