[Bug]ChatterBox 23-Lang Russian model produces gibberish (MPS) / IndexError (CPU) — token/vocab mismatch

[vlenchik]

Environment*

• Platform: macOS, Apple Silicon (M-series)
• ComfyUI: Desktop app, version 0.31.0
• Engine: ChatterBox Official 23-Lang Engine, language = Russian

Summary
Generating Russian speech with the ChatterBox 23-Lang engine produces two different failure modes depending on device setting, both pointing to the same root cause: the Russian model's generated speech tokens exceed the vocabulary size of the s3gen flow decoder's embedding table.

On device=cpu: Hard crash with IndexError during generation.

On device=mps: No crash — generation "succeeds" and returns audio of the expected duration, but the output is unintelligible gibberish (correct prosody/expression, but no real phonetic content in any language). This is consistent with MPS's embedding kernel skipping bounds-checking that CPU enforces — the same out-of-range token issue, just silently corrupting output instead of raising an exception.

Steps to reproduce

1. Add "ChatterBox Official 23-Lang Engine" node, set language to Russian
2. Add "TTS Text" node, connect Engine output
3. Provide a clean reference voice sample (tested with two different samples — a personal recording and a separate professional voice sample — same result both times, ruling out reference audio as the cause)
4. Provide matching Russian reference text transcript
5. Set device=cpu → crash. Set device=mps → completes but output is gibberish.

CPU traceback
IndexError: index out of range in self
File ".../engines/chatterbox/models/s3gen/flow.py", line 218, in inference
token = self.input_embedding(torch.clamp(token, min=0)) * mask
File ".../torch/nn/functional.py", line 2567, in embedding
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
IndexError: index out of range in self

Full traceback available on request — happy to paste more context.

Notes

• Re-downloaded the Russian model files fresh (t3_cfg.safetensors, tokenizer.json, s3gen.safetensors, ve.safetensors, conds.pt) to rule out a corrupted download — same result.
• Ruled out text encoding (confirmed clean UTF-8 Cyrillic) and reference audio format (44.1kHz mono WAV, clean recording) as causes.
• Suspect the Russian language pack's tokenizer/vocab doesn't match the s3gen flow model's embedding table size for this checkpoint.

[diodiogod]

Thanks for the detailed report. We could not reproduce this on our end with the current TTS Audio Suite code.

We tested ChatterBox Official 23-Lang with Russian text on both CPU and CUDA, using both model v2 and v3. Generation completed successfully in every case. We could not test MPS locally, but the reported CPU IndexError did not occur either.

The current Official 23-Lang implementation already bounds-checks and clamps S3Gen token IDs before the embedding lookup. Also, the traceback path in the report points to:

engines/chatterbox/models/s3gen/flow.py

rather than:

engines/chatterbox_official_23lang/models/s3gen/flow.py

That suggests the installation may be outdated, mixed, duplicated, or loading a different engine copy than expected. We do not want to add another clamp blindly because that could hide an installation/model mismatch and merely turn the crash into corrupted audio.

Could you please provide:

• The installed TTS Audio Suite version or commit hash
• The complete traceback, not only the final frames
• The complete ComfyUI startup log and generation log
• The workflow JSON used to reproduce the problem
• The filesystem path from which TTS Audio Suite is loaded
• Confirmation that there are no duplicate TTS Audio Suite installations under custom_nodes
• Results after updating to the current main branch and restarting ComfyUI completely

If the problem still occurs after updating, please include the exact Russian input text, reference transcript, selected model version (v2 or v3), PyTorch version, and macOS version. Those details should let us determine whether this is specifically an MPS/PyTorch issue or an installation/model mismatch.

[diodiogod]

and remember to check if you do not have our old extension, the chatterbox one. It will have conflicts with TTS Audio Suite.