Webpage-to-Speech

Given a URL, this service return an audio file / stream (in WAV format) that reads out the main content of the webpage.

Deployment

To run locally

1. Ensure that you have conda ready. If not, may I suggest Mambaforge?
2. Run synesthesiam/docker-mozillatts: docker run -it -p 5002:5002 synesthesiam/mozillatts.
3. Create a conda env: conda env create -n tts -f conda-requirements.txt -y.
4. Activate the env: conda activate tts.
5. Start the server with Gunicorn: gunicorn main:app --bind 0.0.0.0:80 --timeout 3600 --worker-class sanic.worker.GunicornWorker.
   • I don't use the HTTP server that comes with Sanic, because macOS complains "The process has forked and you cannot use this CoreFoundation functionality safely".
   • As seen in the --timeout option, a request allowed to run for 1h only. Very large text, therefore, may fail.
   • I migrated from Flask to Sanic because Sanic natively supports async view functions, which saves me from interacting with asyncio.get_event_loop().

To run with Docker

Simply do:

docker-compose up --build

Something to note:

• The Dockerfile in this repo is for the URL-to-audio web server only. It still requires the synesthesiam/docker-mozillatts image to be running in a container. Therefore, although you can manually set up the 2 containers, the Docker Compose way is always going to be easier.
• It uses Gunicorn instead of the vanilla Sanic HTTP server.

The containers work together like this:

Sidenote: The diagram above is generated with this command:

docker run --rm -it --name dcv -v $(pwd):/input pmsipilot/docker-compose-viz render -m image docker-compose.yml

Usage

To hear the playback,

• If you are on macOS, ensure that sox is installed: brew install sox. This provides the playback command play.
• If you are on Linux, aplay should do.

Now, you can convert a webpage (using https://sjmulder.nl/en/ as an example) into audio using:

curl -G --output - \
    --data-urlencode 'url=https://sjmulder.nl/en/' \
    'http://localhost:80/' | \
    play -

Of course, you can always save the returned audio as a file and work from there.

To-dos

Here's a list of future features and tasks:

• Investigate why aiohttp calls to the synesthesiam/docker-mozillatts would fail (See the Notebook).
• Make .lrc lyrics or subtitles to go with the audio file.
• Make the speech read out different formats with different audio clues. For example:
  • bolded text can be read with emphasis or notification sounds,
  • read out the numbering of list items, no matter whether the list is ordered or unordered,
  • different levels of headings should be announced ("## lorem ipsum" should sound like "Section 1: lorem ipsum."), and
  • images should be announced with their alt titles. If that's unavailable, call a image labeling service to generate a caption on-the-fly.

License

GPL v3. See LICENSE.