Browser-sized Chatterbox: stub speech encoder + precomputed voice

e52efc2 verified 2 months ago

1.19 kB

license: mit
library_name: transformers.js
pipeline_tag: text-to-speech
base_model: onnx-community/chatterbox-ONNX
tags:
  - tts
  - chatterbox
  - webgpu
  - onnx

Chatterbox for the browser

A repackaging of onnx-community/chatterbox-ONNX sized for in-browser use.

Two changes from upstream:

speech_encoder is a 6 KB stub. transformers.js hardcodes it into Chatterbox's session list, so it is fetched whether or not you use it — 564 MB of model that a text-to-speech app never runs. The speaker conditioning it would produce is instead precomputed once and shipped as voices/default.bin (250 KB), passed straight to generate() as speaker_embeddings, speaker_features, audio_features and audio_tokens.
Only the q4f16 language model is included, not the fp32 copy.

Together these take the download from ~1.4 GB to ~865 MB.

Because the stub is never called, this repo cannot clone a new voice — it can only speak in the bundled one. Use the upstream repo for cloning.

Built for Natural Voice. Weights remain MIT, as upstream.