Samuel

Samuel is a model that learns to control this silly mouth to mimic speech. Say something and it will parrot after you. If you're shy, try one of the pre-made clips or the text-to-speech.

The mouth itself is Pink Trombone, a project originally by Neil Thapen, described by him as "bare-handed speech synthesis".

https://github.com/user-attachments/assets/f952b7d9-43af-4521-8431-0abf9ec027dd

Try it here.

Read my blog post about how it was made here and this one about how I trained the text-to-speech.

How to run it

To just run the code as-is (requires uv and pnpm):

uv run --extra server python -m samuel.server

For development, run these two in separate processes:

# from webapp/
pnpm dev
from repo root
uv run --extra server python -m samuel.server --no-frontend --port 8471

How it was trained

Check out the blog post.

License

Samuel is licensed under the GNU General Public License v3.0 (see LICENSE).

Pink-Trombone/ is a modified version of https://github.com/zakaton/Pink-Trombone/, itself a programmable version of Neil Thapen's original Pink Trombone.

pocket-tts/ is a modified version of Kyutai's Pocket TTS, MIT-licensed (see pocket-tts/LICENSE).

The model is trained on Libri-Light (MIT license), a dataset created from LibriVox audiobooks. LibriVox recordings are in the public domain.