Please note: This code is currently in a very preliminary state, i.e. it would be hard to use out-of-the-box. We hope to clean it and make it more usable in the near future.

The ZeroSpeech challenges aim to answer the question of how we can build speech processing systems directly from speech audio without any labels. It has the dual motivation of understanding language acquisition in humans and developing technology for extremely low-resource languages. The task in ZeroSpeech 2019 is "TTS without T", i.e. text-to-speech without textual input. This is the repository for suzerospeech, the Stellenbosch University ZeroSpeech 2019 system.

The code provided here is not pretty. But we believe that research should be reproducible. We provide no guarantees with the code, but please let us know if you have any problems, find bugs or have general comments.

• docker/
• data/ - Any data files that we produce or get from the challenge organisers.
• features/ - Input features (MFCCs, filterbanks, etc.) are extracted here.
• wavenet/ - WaveNet speech synthesis.
• notebooks/
  • vq_vae.ipynb
  • cat_vae.ipynb
• evaluation/
• src/ - Mature source used in different parts of the project can be put here.

This recipe comes with Dockerfiles which can be used to build images containing all of the required dependencies. This recipe can be completed without using Docker, but using the image makes it easier to resolve dependencies. At the moment, we use a Dockerfile which is different from the Dockerfile provided as part of the challenge. To use our docker image you need to first:

• Install Docker and follow the post installation steps.
• Install nvidia-docker.

To build the docker image, run the following:

cd docker
docker build -f Dockerfile.tf-py36.cpu -t tf-py36 .
cd ..

There is also a GPU version of the image. The rest of the steps in this recipe can be run in a container in interactive mode. Start the docker image with the required data directories mounted:

docker run \
    -v ~/endgame/datasets/zerospeech2019/shared/databases/english/:/data/english \
    -v "$(pwd)":/home -it -p 8887:8887 tf-py36

To run on a GPU, --runtime=nvidia is additionally required.

To directly start a Jupyter notebook in a container, run:

docker run --rm -it -p 8889:8889 \
    -v ~/endgame/datasets/zerospeech2019/shared/databases/english/:/data/english \
    -v "$(pwd)":/home \
    tf-py36 \
    bash -c "ipython notebook --no-browser --ip=0.0.0.0 --allow-root --port=8889"

and then open http://localhost:8889/ in a browser.

If you are not using the docker image, install all the standalone dependencies (see Dependencies section below). Then follow the steps here. The docker image includes all these dependencies and GitHub repositories.

Clone the required GitHub repositories into ../src/ as follows:

mkdir ../src/  # not necessary using docker
cd ../src/
git clone https://github.com/jameslyons/python_speech_features
cd python_speech_features
python setup.py develop
cd ../../suzerospeech2019/

Move to features/ and execute the steps in features/readme.md.

• British spelling for naming and documentation.
• Use double quotes "..." for Python strings.

• Python 3
• python_speech_features

This code is distributed under the Creative Commons Attribution-ShareAlike license (CC BY-SA 4.0).