cadia-lvl / unit-selection-festival

First version of Unit Selection recipe for Icelandic from Reykjavík University

Unit selection recipe for Icelandic

This is a default unit selection recipe for The Festival Speech Synthesis System. The voice configuration inside voice/ was built using $FESTVOX/src/unitsel/setup_clunits and then adapted to work for Icelandic data.

Table of Contents

• Installation
• Training
  • Training with Docker
• Running
  • Running with Docker
• License
• Authors/Credit
  • Acknowledgements

Installation

This recipe is built with The Festival Speech Synthesis System. To train and run a voice model you will need to install:

• The Festival Speech Synthesis System
• The Edinburgh Speech Tools Library
• Festvox
• fairseq g2p models

Training

Training a voice requires a voice corpus, a fairseq grapheme to phoneme model and an optional lexicon. This documentaion assumes you are using the Talromur corpus and you have the data and a g2p model available in a directory called ext/.

ext/ directory format:

data/audio/*.wav
data/index.tsv
g2p-lstm/

You can edit the run script to adapt it to your own data.

If you are using the Talrómur corpus and have set up the ext/ directory correctly training the voice is as simple as:

cd voice
./run.sh ../ext/data standard

If you want to use your own data replace the parameters to the run script with your data folder and your g2p model.

Training with Docker

If you wishi, you can train your voice in a Docker container. To do this simply build a container using the included configuration in train.Dockerfile and run it:

docker build . --tag lvl-us-is-train -f train.Dockerfile
docker run -v ${PWD}/ext/data:/usr/local/src/ext/data -v ${PWD}/ext/g2p-lstm/:/app/fairseq_g2p -v ${PWD}/voice/:/usr/local/src/voice lvl-us-is-train:latest

Running

After training a model you can synthesize new audio. The synthesis process does need the same grapheme to phoneme model as used for training in addition to the files and folders produced in the training process.

# Synthesize by calling something like this:
# This only works within the voice directory
echo "Halló, ég kann að tala íslensku." | python3 normalize.py - - | ../festival/bin/text2wave -eval festvox/lvl_is_v0_clunits.scm -eval '(voice_lvl_is_v0_clunits)' > demo.wav

Running with Docker

If you are using Docker you can synthesize by building and running runtime.Dockerfile.

docker build . --tag lvl-us-is-run -f runtime.Dockerfile
# The docker container uses the default input file: /opt/voice/input.txt
docker run lvl-us-is-run:latest > output.wav

# You can change input.txt to any file with a single normalized utterance
# If you want to change the input text, do something like this command
docker run -v ${PWD}/hvad_segir_thu.txt:/opt/voice/input.txt lvl-us-is-run:latest > hvad.wav

License

This recipe is published under the MIT license.

Authors/Credit

Reykjavik University

Þorsteinn Daði Gunnarsson [e-mail]

Acknowledgements

This project was funded by the Language Technology Programme for Icelandic 2019-2023. The programme, which is managed and coordinated by Almannarómur, is funded by the Icelandic Ministry of Education, Science and Culture.