chatterbox-tts-onnx

Chatterbox — ONNX export for Vernacula

A complete ONNX export of ResembleAI/chatterbox (Resemble AI's open zero-shot TTS: an 0.5B Llama-style speech-token LM over a CosyVoice-style flow-matching vocoder), for use as the Chatterbox text-to-speech engine in Vernacula.

• Conversion script: scripts/chatterbox_export/
• Vernacula: github.com/christopherthompson81/vernacula
• Upstream model: ResembleAI/chatterbox

Highlights

• All four stages in one reproducible recipe, pinned to an upstream revision. The other ONNX sources either publish artifacts without the export (Resemble), a bundle assembled from several places (onnx-community), or three of the four graphs without the language model.
• Language model with KV cache, exported from scratch, so a rollout is one graph call per step.
• Vocoder in both layouts: a single merged graph that runs the whole CFM solve as an ONNX Loop (what the C# runtime loads), and the three split graphs it was built from (flow encoder, CFM estimator, mel-to-wave) for orchestrating the Euler loop outside the graph. Per-graph parity against upstream eager passes on all five checks.
• Word timings for free: the LM's cross-attention over the text tokens gives per-word alignment during synthesis, no forced-alignment pass — see docs/tts_alignment_desync_investigation.md.

Contents

Every graph keeps its weights in an external-data sidecar named <graph>.onnx_data (that spelling, with an underscore, is what the graphs reference — keep the pair together).

The bundle is about 9.3 GB in fp32.

Export provenance

Exported via scripts/chatterbox_export/ in the Vernacula repo (export_chatterbox_to_onnx.py --device cuda --dtype float32). The export applies scoped, self-restoring patches to the upstream modules for tracing — real-valued STFT/iSTFT in place of the complex ones, a dynamic-shape solve_euler, deterministic noise — and records the pinned ResembleAI/chatterbox revision it read in export-report.json. The design log is docs/chatterbox_investigation.md; the merged-Loop vocoder is Run 10 of docs/chatterbox_perf_investigation.md.

License

MIT, inherited from ResembleAI/chatterbox. tokenizer.json is upstream's file, redistributed unchanged.

Not included: upstream's Perth audio watermarker. The Python package applies it to every output; this export is the synthesis path only. Please respect the upstream project's guidance on responsible use of voice cloning.

Using these files

In Vernacula, point Settings → Text-to-Speech → Chatterbox at a folder holding these files (or use its Download button) and pick a reference clip per job. Outside Vernacula the graphs need orchestration — speaker encoding, the LM rollout with its KV cache, then the vocoder — which the C# ChatterboxPipeline in src/Vernacula.Tts.Base implements end to end; it is the reference for the I/O names and shapes.

from huggingface_hub import snapshot_download
import onnxruntime as ort

path = snapshot_download(repo_id="christopherthompson81/chatterbox-tts-onnx")
lm = ort.InferenceSession(f"{path}/language_model.onnx")
print([i.name for i in lm.get_inputs()])   # token embeddings + KV cache in; logits + KV cache out

Limitations

English only. A rollout is capped at 1024 LM steps per chunk; Vernacula splits documents into paragraphs first and flags a chunk that hits the cap. Everything else is upstream's — see the Chatterbox model card.

Citation

See the upstream model card and github.com/resemble-ai/chatterbox.

Acknowledgments

• Original model: Resemble AI (Chatterbox)
• ONNX export: Chris Thompson for Vernacula

Issues with the ONNX export specifically: open an issue on the Vernacula repo. Issues with the underlying model: see the upstream model card.

See also

• Vernacula on GitHub — the speech pipeline app this package is built for
• Conversion script (scripts/chatterbox_export/)
• ResembleAI/chatterbox — upstream model card
• Other Vernacula model packages