Chatterbox Nano: ONNX for the browser (WebGPU)
An ONNX Runtime Web export of ResembleAI/chatterbox-nano
(revision 71ccd1d0081b430592cea481f4307e764e07bc64), used by StudyWhile to make flashcard audio
on a laptop's GPU in the browser.
| File | What it is |
|---|---|
t3.onnx |
T3 speech-token model: one graph for prefill and decode (variable row count) |
encoder.onnx, cfm.onnx, f0.onnx, vocoder.onnx |
S3Gen MeanFlow decoder, flow and vocoder (variable sizes) |
tables.safetensors |
Text and speech token embeddings and position table |
tokenizer/ |
GPT-2 tokenizer files from the base model |
Weights are stored as fp16 with casts, so computation runs in fp32 on WebGPU. No voice, speaker conditioning or reference audio is included: the speaker conditioning is an input supplied at run time.
License
MIT, the same as the base model. Copyright (c) 2025 Resemble AI. See LICENSE.
Model tree for serendipbrity/chatterbox-nano-onnx-webgpu
Base model
ResembleAI/chatterbox-nano