Build scripts for Docker images of the dnn/hmm grapheme based speech synthesizers

Create docker images for an Ossian based speech synthesizer.

Build images

setup.sh is building two images for each (dnn/hmm):

• full Ossian synthesizer with training support for the new voices
• synthesizer only

To build DNN images without licence restrictions use:

./setup.sh dnn

and

./setup.sh hmm

To build commercially restricted conda versions use:

./setup-with-conda.sh dnn

and

./setup-with-conda.sh hmm

The build scripts assume that you have

• Ossian registration credentials (if needed) in setup-ossian.sh (see Ossian demo)
• Installed podman and buildah: 'sudo apt-get install podman buildah'
• if needed add for your user:
  • in /etc/subuid

    yourusername:100000:65536
    

  • in /etc/subgid

    yourusername:100000:65536
    

  Replace yourusername whit your linux username.

Train a tiny voice

test-train.sh will use the image dnn-greem-train to train a sample voice. test-train-hmm.sh will use the image hmm-grafeem-train to train a sample voice.

Conda versions can be tested by adding conda parameter:

./test-train.sh conda

Also the image dnn-grafeem-train is available at ghcr.io/jaotus/dnn-grafeem-train:latest and can be tested with test-synthesize-cloud.sh.

Customization

Synthesizer parameters can be customized by modifying the files

• train.sh
• train-hmm.sh
• synthesize.sh
• synthesize-hmm.sh

Container images allow mapping of the following folders:

• /Ossian/voices
• /Ossian/corpus
• /Ossian/train
• /Ossian/test

Also some Ossian parameters can be specified for these scripts:

LANGUAGE=et
SPEAKER=sml_tonu

Windows and Docker Desktop

These containers can also be run on Windows10/11 systems. For this

• install Docker desktop (see Windows installation)
• when asked select WSL 2 backend and Ubuntu Linux (default)

Now you can start Ubuntu and verify that command docker prints the help message.

Training

The test-* scripts work with docker and podman also.