Building Voices in Festival
       Processes and issues in building speech synthesis voices
                      festvox-2.1-release
                 Alan W Black and Kevin A. Lenzo
                         awb@cs.cmu.edu
                     http://www.festvox.org

This is to announce a new release of the festvox project.  The festvox
project, based at Carnegie Mellon University, distributes
documentation, scripts and examples that 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.

The release includes:
 o Support for designing, recording and autolabelling diphone databases
 o Support for designing, recording and autolabelling unit selection databases
 o Support for designing, recording and autolabelling clustergen parametric
     voices.
 o Support for Nagoya Institute of Technologies HTS synthesis system.
 o Building simple limited domain synthesis engines
 o Support for building rule driven and data driven prosody models
 o Lexicon and building letter to sound rule support
 o Predefined scripts for building new US (and UK) English voices
 o Example diphone and limited domain synthesis databases

Since the last full release (Jan 2003) apart from general bug
fixes and improvements this release includes
 o Better clunits general voice support
 o Clustergen Statistical Parametric Synthesis (HTS like) that is
     easier to use and relaible on multiple voices and languages
     http://www.festvox.org/bsv/c3170.html     
 o A new automatic phoneme label EHMM, which is easier to use and
   give better results that our previous methods.
      festvox/src/ehmm/
 o VC voice conversion module (as standalone and integrated into Festival)
      festvox/src/vc/
 o Support for finding "nice" prompts in large databases of texts in
   new languages.
      http://www.festvox.org/bsv/c2174.html
 o Full support under Cygwin under windows for voice building

The complete documentation in html, and in downloadable format including
the scripts and programs necessary to build new voices and example
databases is available from

        http://www.festvox.org/bsv/

The full distribution is packaged, including postscript and the
generated html, and is available from

        http://www.festvox.org/festvox/festvox-2.1-release.tar.gz

LICENCE

This documentation and related scripts is free software, distributed
under an X11-type licence (like Festival itself).  No claims are made
by the authors of this work, Carnegie Mellon University (or the
University of Edinburgh), on the voices that you generate with the
scripts and techniques described within this distribution.

REQUIREMENTS

A Unix Machine (Linux. FreeBSD, Solaris etc) with working audio i/o:
    although there is nothing inheritantly Unix about the scripts.
Or a Windows machine running the cygwin environment.
Edinburgh University's Festival Speech Synthesis System and 
The Edinburgh 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
       http://www.festvox.org/festival/latest/

    Note you must use Speech Tools 1.2.96 and Festival 1.96 or later.

    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 festvox.
EMU Labeller
    The University of Macquarie's Speech Hearing and Language Research
    Centre distribute labelling tools for speech databases.  We use it
    here for viewing speech, as spectrograms, F0 contours, phone labels etc.
    It is available from 
           http://www.shlrc.mq.edu.au/emu/
    Other waveform labeller/viewers exist and you find them more convinient
    to use but we include support for emulabel as it meets our requirements
    and is freely available.
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.

WARNING

This is not a 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.

However, this release does include complete simple walkthroughs of
scripts that can build voices in English with little more than
recording time, by people without knowledge of scheme programming or
speech technology, but the results will be better if you take time
to understand the underlying processes.

Also note there are still unwritten parts of the documentation,
new releases in the future with reduce such parts.

INSTALL

    download http://www.festvox.org/festvox/festvox-2.1-release.tar.gz
    unpack its and see festvox/README for instructions for installation
    and use.

-------
Alan and Kevin
Pittsburgh, PA
21st January 2007