vando2999/editorbr-chatterbox-ptbr

Chatterbox Multilingual pt-BR: ONNX

Resemble AI's Brazilian Portuguese finetune of Chatterbox Multilingual V3 (ResembleAI/Chatterbox-Multilingual-pt-br, t3_pt_br.safetensors, sha256 074aaf65…cbb6c), put into the graphs of the community ONNX export Masterx/chatterbox-multilingual-v3-ONNX at revision 19b94afe41de738624b2f3afc75b4cd56221fbeb.

This is what Editorbr downloads to read Brazilian Portuguese in a cloned voice, on the user's own machine.

What is here

The three graphs that hold T3 weights - embed_tokens, speech_encoder (its conditioning encoder) and language_model (the full-precision Llama backbone) - with every T3 tensor replaced by the finetune's. The graph files themselves are byte-identical to the base export; only their _data changed.

Nothing else differs from the base export, so take these from there: tokenizer.json, onnx/conditional_decoder.onnx and its _data (S3Gen is the shared V3 decoder).

How it was made

Each float initializer of the base graphs was matched, by the hash of its bytes, to a tensor of the base checkpoint t3_mtl23ls_v3.safetensors - as it is, transposed, or fused as onnxruntime-genai lays out a layer's q/k/v and gate/up projections - and replaced by the same arrangement of the finetune's tensor. 212 backbone tensors, 5 embedding tensors and 14 conditioning-encoder tensors were swapped; every T3 tensor the finetune changed is covered except text_head, which inference does not use. The S3 tokenizer, CAMPPlus and the rotary caches are untouched.

Checked end to end through Editorbr's Rust runtime (greedy decoding, repetition penalty 1.2, the base export's default_voice.wav): Whisper small transcribes the read back word for word.

Licence

MIT, as Chatterbox. All credit for the model goes to Resemble AI; this is a format conversion. The ONNX graphs carry no Perth watermarker - Resemble asks that generated audio be used responsibly.