FestCat Voice Builder

This set of Makefiles and templates have been written to automate the building of HTS voices in:

• Catalan for the FestCat project
• English, with the ARCTIC dataset

This Makefiles and templates can be extended to support building of HTS voices in other languages.

Quick Start

1. Install dependencies. You will need some general utilities:

   • C, C++ compilers
   • make
   • autoconf
   • grep, awk, sed
   • sox
   • perl
   • csh
   • I believe I also needed the Debian package libx11-dev to build HTS and maybe? tcl-snack. Sorry for not providing more details.

2. Register and agree to the HTK license at http://htk.eng.cam.ac.uk/ You need the HTK user name and password so HTK can be downloaded.

3. Train a voice. The training can take time, GB of disk and a fair amount of RAM. For instance, training the ona Catalan voice may take several days, <10 GB of hard disk space with peaks of >8??GB of RAM. In order to train HTS voices, run:

    ./configure htk_user="yourhtkuser" htk_password="yourhtkpassword"
    make ca_ona.spk
    make ca_bet.spk
    make ca_pau.spk
    make en_slt.spk
    make en_awb.spk
    make fr_zoe.spk
   

Dependencies

Language independent tools automatically downloaded, built and run

The following software is not language specific and is downloaded to the deps directory, compiled and installed automatically to the tools directory.

1. Speech Tools 2.4-release
2. Festival 2.4-release
3. SPTK 3.10
4. HTS-2.3.1 for HTK-3.4.1 (downloading HTK requires a user and password, as well as agreeing to the non-free license)
5. HDecode 3.4.1 (it has similar restrictions to HTK)
6. hts_engine 1.10

Language specific tools

Any language specific tool should be installed automatically from the Makefile available in data/$lang.

Catalan

This is downloaded automatically:

• upc_ca_base (also known as festival-ca in some GNU/Linux distributions)
• raw recordings and utt files.

English

This is downloaded automatically:

• festlex_CMU and festlex_POSLEX
• recordings and utt files from ARCTIC dataset.

Usage

Training Catalan voices may take several days, some GB of Hard disk and a lot of RAM.

In order to train HTS voices, run:

./configure htk_user="yourhtkuser" htk_password="yourhtkpassword"
make ca_ona.spk
make ca_bet.spk
make ca_pau.spk
make en_slt.spk
make en_awb.spk