"Building Voices in Festival"
 Alan W Black (awb@cs.cmu.edu), Kevin Lenzo (lenzo@duolingo.com)
                 and see ACKNOWLEDGEMENTS
                  http://www.festvox.org

For full details about voice building see the document itself

http://festvox.org/bsv/

The included documentation, scripts and examples should be sufficient for an interested person to build their own synthetic voices in currently supported languages or new languages in the University of Edinburgh's Festival Speech Synthesis System. The quality of the result depends much on the time and skill of the builder. For English it may be possible to build a new voice in a couple of days work, a new language may take months or years to build. It should be noted that even the best voices in Festival (or any other speech synthesis system for that matter) are still nowhere near perfect quality.

This distribution includes

Support for designing, recording and autolabelling statistical parametric
    synthesis voices
Support for designing, recording and autolabelling diphone databases
Support for designing, recording and autolabelling unit selection dbs
Building simple limited domain synthesis engines
Support for building rule driven and data driven prosody models
   (duration, intonation and phrasing)
Support for building rule driven and data driven text analysis
Lexicon and building Letter to Sound rule support
Predefined scripts for building new US (and UK) English voices
Predefined scripts for building grapheme(++) voices for any language
Scripts for designing and selecting prompts to record for
   arbitrary languages

ROADMAP FOR 3.0: GENERATIVE AI FOCUS

Version 3.0 will represent a paradigm shift towards modern generative AI-based speech synthesis, while maintaining backward compatibility with traditional voice building approaches. The focus will be on:

Core Generative AI Models:

Advanced Neural Architectures:
    - Full VITS (Variational Inference with adversarial learning for TTS)
    - Diffusion-based TTS models 
    - Large Language Model integration for text analysis

Modern Infrastructure:
    - GPU-accelerated training and inference
    - Distributed training support
    - Cloud deployment ready architectures
    - ONNX model export for cross-platform deployment
    - WebAssembly support for browser-based synthesis

Developer Experience:
    - Python-first API with simplified voice building workflows
    - Jupyter notebook tutorials and examples
    - Pre-trained model zoo with easy fine-tuning
    - Automatic hyperparameter optimization
    - MLOps integration (experiment tracking, model versioning)

The 3.0 release will maintain the robust voice building foundation of Festvox while embracing the latest advances in generative AI, making high-quality voice synthesis accessible to a broader community of researchers and developers.

New in 2.9

FALCON: Neural extension to festvox voice building suite
Support for neural TTS models including Tacotron, WaveNet, and WaveRNN
VITS-lite implementation for end-to-end neural synthesis
Modular neural architecture with hierarchical blocks (Layers -> Blocks -> Models)
Integration with PyTorch for modern deep learning workflows
Support for neural vocoders (WaveGlow, WaveRNN)
Phoneme-level and acoustic modeling with neural networks
Judith framework integration for automated experimentation

New in 2.8

https://github.com/festvox/festival/
Grapheme built voices can be converted to .flitevox files for android
Database size reduction for random forest clustergen voices
Random Forests for F0 prediction too
18 English voices, and 13 Indic voices

New in 2.7

Random forest models building for spectrum and duration in clustergen
Grapheme based synthesizers (with specific support for large number
  of unicode writing systems)
Clustergen state and stop value optimization
Wavesurfer label support
SPAM F0 support
Phrase break support
Support for SPTK's mgc parameterization

New in 2.3

Support for cygwin tools under Windows
Substantially improved CLUSTERGEN support with mlpg and mlsb

WARNING

This is not a pointy/clicky plug and play program to build new voices. It is instructions with discussion on the problems and an attempt to document the expertise we have gained in building other voices. Although we have tried to automate the task as much as possible this is no substitute for careful correction and understanding of the processes involved. There are significant pointers into the literature throughout the document that allow for more detailed study and further reading.

REQUIREMENTS

A Unix Machine

although there is nothing inheritantly Unix about the scripts, no
attempt has yet been made about porting this to other platforms

Festival and Speech Tools

This uses speech tools programs and festival itself at various
stages in builidng voices as well as (of course) for the final
voices.  Festival and the Edinburgh Speech Tools are available from

   http://www.cstr.ed.ac.uk/projects/festival/
   
or

   http://www.festvox.org/festival

or

   https://github.com/festvox
   
It is recommended that you compile your own versions of these
as you will need the libraries and include files to build some
programs in this festvox.

Wavesurfer

To display waveforms, spectragrams and phoneme labels.

Patience and understanding

Building a new voice is a lot of work, and something will probably
go wrong which may require the repetition of some long boring and
tedious process.  Even with lots of care a new voice still might 
just not work.  In distributing this document we hope to increase the
basic knowledge of synthesis out there and hopefully find people 
who can improve on this making the processing easier and more reliable
in the future.

INSTALLATION

You must have the Edinburgh Speech Tools and Festival instllation before you can build the tools in the festvox distribution.

Unpack festvox-2.8-release.tar.gz or clone it from github

git clone https://github.com/festvox/festvox
cd festvox
./configure
make

The configuration basically tries to find your version of the Edinburgh Speech Tools and uses its configuration to set compiler type etc. So you must have that installed. If configure fails try expliciting setting your ESTDIR environment variable to point ot your compiled version of the Speech Tools.

A pre-generated version of the document in html and postscript are distributed in the html/ directory

If you need to build the document itself, you will need a working version of the docbook tools, which may (or may not) already be installed on your system

To build the documenation

cd docbook
make doc

Note that even if the documentation build fails you can still use all the scripts and programs.

To use the scripts and programs in the festvox distribution each user is expected to have the environment variables ESTDIR and FESTVOXDIR set for example as (if you use bash, zsh, ksh or sh)

export ESTDIR=/home/awb/projects/speech_tools
export FESTVOXDIR=/home/awb/projects/festvox
export FLITEDIR=/home/awb/projects/flite
export SPTKDIR=/home/awb/projects/SPTK

Or if you use csh or tcsh

setenv ESTDIR /home/awb/projects/speech_tools
setenv FESTVOXDIR /home/awb/projects/festvox
setenv FLITEDIR /home/awb/projects/flite
setenv SPTKDIR /home/awb/projects/SPTK

Remember to set these to where your installations are, not ours.