Home

Awesome

NOTE: Please see coqui-docker for docker images of Coqui TTS (Mozilla TTS's successor)

Mozilla TTS

Multi-platform Docker images for Mozilla TTS. Many thanks to erogol and the community!

Screenshot of web interface

Supported languages (see Released Models):

Supported platforms:

RAM Limitations

If you're running on a Raspberry Pi with only 1 GB of RAM, you may be unable to load some of the larger models without increasing your swap space. To do this, simply edit the /etc/dphys-swapfile file (with sudo) and increase CONF_SWAPSIZE (1000 is recommended, value is MB). Make sure to reboot after editing this file.

Using

$ docker run -it -p 5002:5002 synesthesiam/mozillatts:<LANGUAGE>

where <LANGUAGE> is one of the supported languages (en, es, fr, de). If no language is given, U.S. English is used.

Visit http://localhost:5002 for web interface.

Do an HTTP GET at http://localhost:5002/api/tts?text=your%20sentence to get WAV audio back:

$ curl -G --output - \
    --data-urlencode 'text=Welcome to the world of speech synthesis!' \
    'http://localhost:5002/api/tts' | \
    aplay

HTTP POST is also supported:

$ curl -X POST -H 'Content-Type: text/plain' --output - \
    --data 'Welcome to the world of speech synthesis!' \
    'http://localhost:5002/api/tts' | \
    aplay

A /process endpoint is available for compatibility with MaryTTS. Expose the correct port (59125) for maximum compatibility:

$ docker run -it -p 59125:5002 synesthesiam/mozillatts

You should now be able to use software like the Home Assistant MaryTTS integration. Note that only the INPUT_TEXT field is actually used.

Custom Model

The Docker image is usually built with buildx for multi-platform support. If you just want to build an image for one platform, you can do this:

$ NOBUILDX=1 LANGUAGE=en scripts/build-docker.sh

When you set a LANGUAGE, the build script looks in model/<LANGUAGE>. These files should exist:

Optionally, you may also include a vocoder:

If the sample rates between the model and vocoder don't match, the audio will be interpolated.

Docker Download Cache

When building the Docker image, the download directory may contain architecture-specific Python wheels. The download/amd64 directory, for example, will be used with pip's --find-links on x86_64 systems.

The download/shared directory is used for all architectures. If a requirements.txt file is present there, it is used to install dependencies for MozillaTTS. This can be used to exclude Tensorflow, etc., or to use specific package versions.

Use Docker buildx

To use buildx, you'll need to enable experimental features in the Docker CLI and then set up a private registry:

$ docker run -d -p 15555:5000 --name registry --restart=always registry:2

This registry runs on port 15555. Next, create a configuration file at /etc/docker/buildx.conf with this inside:

[registry."localhost:15555"]
  http = true
  insecure = true

Note the same port number (15555). Finally, run the following commands to create a builder:

$ docker run --rm --privileged multiarch/qemu-user-static --reset -p yes
$ docker buildx create --config /etc/docker/buildx.conf --use --name mybuilder
$ docker buildx use mybuilder
$ docker buildx inspect --bootstrap

For some reason, these have to be run again after every reboot and will sometimes require removing the builder first.

If all is well, you can build for specific platforms like this:

$ PLATFORMS=linux/arm/v7 LANGUAGE=en DOCKER_REGISTRY=localhost:15555 scripts/build-docker.sh

Note that the limiting factor for most platforms is a compiled PyTorch wheel. Pre-built wheels are available here for ARM and PyTorch 1.6.0. Put wheels in the download directory before building.