videopython-chatterbox

Chatterbox TTS by Resemble AI, republished with corrected dependency metadata.

Version 0.1.7.post2 also fixes short-text alignment: repetition detection now checks that earlier text tokens exist before reducing their attention scores. Inputs with five or fewer text tokens retain long-tail and token-repetition EOS checks without raising an empty-reduction error. The distribution-name lookup and corrected dependency metadata remain the other changes from upstream 0.1.7.

pip install videopython-chatterbox

The import name is unchanged, so this is a drop-in replacement:

from chatterbox.mtl_tts import ChatterboxMultilingualTTS

Conflict warning: this package and chatterbox-tts both install a top-level chatterbox package. Install one or the other, never both.

Why this fork exists

Upstream chatterbox-tts 0.1.7 pins diffusers==0.29.0, torch==2.6.0, torchaudio==2.6.0, transformers==5.2.0 and safetensors==0.5.3. Those == pins make it unsatisfiable alongside any modern ML stack — for example pyannote-audio>=4.0 requires torch>=2.8.0, so the two cannot coexist:

$ pip install chatterbox-tts diffusers>=0.35.0
  × No solution found: chatterbox-tts==0.1.7 depends on diffusers==0.29.0

The pins are not load-bearing. Chatterbox touches diffusers in exactly one place — its vendored Matcha-TTS decoder — importing 9 symbols (GEGLU, GELU, AdaLayerNorm, AdaLayerNormZero, ApproximateGELU, Attention, LoRACompatibleLinear, maybe_allow_in_graph, get_activation). All of them still exist and behave identically in diffusers 0.39.

Verification

Real synthesis and voice cloning were run on torch 2.13.0 / diffusers 0.39.0 / transformers 5.14.1 / numpy 2.4.6, then repeated under upstream's pinned transformers==5.2.0, with fixed seeds:

Output is bit-identical. Generations were also transcribed with Whisper to confirm the audio matches the requested text, rather than merely being non-silent.

Metadata changes

Two of these are deliberate constraints rather than relaxations:

diffusers<1.0 — LoRACompatibleLinear is deprecated with removal scheduled for diffusers 1.0.0. It is the one symbol that will genuinely break, so the cap is real.

setuptools<81 — resemble-perth declares no dependencies at all and imports pkg_resources, which setuptools removed in 82+. Its __init__.py swallows the resulting ImportError and sets PerthImplicitWatermarker = None, so chatterbox fails later at model load with a confusing TypeError: 'NoneType' object is not callable. Upstream is affected by this too; it is simply undeclared there.

Relationship to upstream

Changes are limited to packaging compatibility and the documented short-text alignment crash fix. If upstream incorporates these fixes and relaxes its pins, prefer chatterbox-tts.

To rebase onto a new upstream release: replace src/chatterbox/ wholesale, re-apply the distribution-name lookup in chatterbox/__init__.py, and check whether the short-text alignment fix is already present before carrying it forward. Run python -m pytest tests/test_short_alignment.py, verify the remaining source against the new sdist, bump version to <upstream>.postN, and re-check the dependency table above.

For 0.1.7.post2, all seven direct PyTorch tests pass (text widths 1–6 and 20). GPU synthesis of “za”, “Ja” and a full Polish sentence also succeeded without videopython's alignment monkeypatch. These checks cover the short-input crash; they do not establish general speech quality or performance.

License

MIT, Copyright (c) 2025 Resemble AI — preserved verbatim in LICENSE. See NOTICE for the fork's provenance statement. Model weights are not redistributed; they are fetched at runtime from the Hugging Face Hub.