Phoneme Cue Baker — TTS model pack (tts_v1)
ONNX exports of the Chatterbox Multilingual TTS model and its supporting voice/feature models, packaged for the offline Phoneme Cue / FaceCue baker. The baker uses these graphs to synthesize missing audio from a transcript and to create a voice bundle from a reference clip, entirely on-device via ONNX Runtime (CPU or DirectML/CUDA).
Graphs are fp32. (They are hosted here rather than on a 2 GB-capped host, so no fp16 conversion is applied.)
Contents (tts_v1/)
| File | Role |
|---|---|
t3_decoder_kv.onnx |
T3 text→speech-token decoder (KV-cache) |
flow_to_mu.onnx |
S3Gen flow front (tokens + prompt → CFM conditioning) |
cfm_estimator.onnx |
CFM ODE step network |
cfm_euler.onnx |
per-step Euler + guidance combine |
hift_vocoder.onnx |
HiFT vocoder (mel → waveform) |
voice_encoder.onnx |
speaker-embedding encoder |
campplus_xvector.onnx |
CAMPPlus speaker x-vector |
s3tokenizer_quantize.onnx |
S3 speech tokenizer |
t3_frontend.bin / .json |
T3 embedding/Perceiver weights |
grapheme_mtl.json |
grapheme tokenizer vocab |
mel_*.npy |
creator mel filterbanks (s3gen / VE / s3tok / Kaldi) |
builtin_voice.fcv1 |
the built-in voice bundle |
A tts.manifest.json lists each file's URL, SHA256, and size for download + integrity verification.
License & attribution
Distributed under the MIT License (see LICENSE), inherited from Chatterbox (Copyright © 2025
Resemble AI). The pack bundles components under Apache 2.0 (the CosyVoice-lineage S3Gen flow,
S3 tokenizer, CAMPPlus, and ESPnet/WeNet conformer building blocks — Copyright © Alibaba Inc and
others) and MIT (the Real-Time-Voice-Cloning voice encoder). Full attribution is in NOTICE.
This pack is a redistribution of upstream work in ONNX form and is not affiliated with or endorsed by Resemble AI, Alibaba, or the other upstream authors.