Chatterbox Flash

Made with ♥️ by

Prior-calibrated block-diffusion zero-shot TTS, extending Chatterbox-TTS with a parallel masked decoder while preserving streaming generation.

Chatterbox-Flash replaces the autoregressive T3 decoder of Chatterbox-TTS with a Fast-dLLM v2-style block-diffusion decoder, adds two inference-time techniques described in our paper — prior-calibrated PMI scoring and early decoding via a time-shifted quantile schedule — and reuses the original S3Gen flow-matching vocoder, Metavoice voice encoder, and English tokenizer unchanged.

The released package is inference-only. Training code is intentionally omitted; all reproductions of the paper's numbers can be driven from the scripts in this repository given the released checkpoints.

Installation

Recommended: uv

First create and activate a virtual environment:

uv venv                       # creates .venv with a compatible Python
source .venv/bin/activate     # Windows: .venv\Scripts\activate

chatterbox-tts==0.1.7 (our base dependency) pins torch==2.6.0, but most modern CUDA wheels in the stack (torchvision, xformers, flash-attn, flashinfer-python) only ship matched ABI binaries for torch 2.7.x. Our pyproject.toml declares a [tool.uv] override-dependencies section that tells uv to honour the 2.7.x pin and skip the upstream 2.6 constraint, so a single command produces a consistent environment:

# Core — uv reads tool.uv.override-dependencies automatically:
uv pip install chatterbox-flash

# Optional — high-throughput inference backend (FlashInfer, CUDA):
uv pip install "chatterbox-flash[flashinfer]"

# Optional — Apple Silicon native Metal backend (mlx + mlx-lm, macOS only):
uv pip install "chatterbox-flash[mlx]"

# Optional — full evaluation suite (SIM-o / WER / UTMOS via OmniVoice):
uv pip install "chatterbox-flash[eval]"

If uv ever fails on a torchvision ABI mismatch (RuntimeError: operator torchvision::nms does not exist) or a flash_attn_2_cuda undefined symbol, force a torch-2.7-matched torchvision:

uv pip install 'torchvision>=0.22,<0.23'

Alternative: plain pip

pip does not understand [tool.uv], so you need to install with --upgrade torch torchaudio after the initial resolve to undo the chatterbox-tts 2.6 pin manually:

pip install chatterbox-flash
pip install --upgrade 'torch>=2.7,<2.8' 'torchaudio>=2.7,<2.8' 'torchvision>=0.22,<0.23'

Local install from source (development)

Clone the repository and install in editable mode so source edits take effect without reinstalling. Pick the extras for your hardware:

git clone https://github.com/resemble-ai/chatterbox-flash.git
cd chatterbox-flash

# CUDA box (FlashInfer + eval):
uv pip install -e ".[flashinfer,eval]"

# Apple Silicon (Metal backend):
uv pip install -e ".[mlx]"

# Minimal (CPU / torch SDPA only):
uv pip install -e .

Engine selection at runtime

Pick the hardware path via --backend on synthesize.py — see the table in Quick start. From the Python API, pass the lower-level engine name (flashinfer / torch / mlx):

from chatterbox_flash import FLASHINFER_AVAILABLE, MLX_AVAILABLE, ChatterboxFlashTTS

tts = ChatterboxFlashTTS.from_pretrained("ResembleAI/chatterbox-flash", device="cuda")
wav = tts.generate(text, audio_prompt_path="ref.wav", backend="torch")

CHATTERBOX_FLASH_ENGINE={flashinfer,torch,mlx} forces an engine per process (handy in CI/CD).

Quick start

A single entry-point synthesize.py covers all four hardware paths via one --backend flag. One reference voice + one or many texts (a single text is just a batch of one).

--backend	Engine	Device	dtype	CUDA graph	Notes
gpu (default)	torch SDPA	cuda	bf16	on	No JIT cold start
flashinfer	FlashInfer	cuda	bf16	on	Paged KV; warmup-amortised throughput
cpu	torch SDPA	cpu	fp16	off	CPU-only validation / Docker (--dtype fp32 to fall back)
mlx	mlx (Metal)	cpu*	fp16	off	Apple Silicon native ([mlx] extra)

* MLX runs the LLaMA backbone on Metal; the PyTorch side stays on CPU.

Override the per-backend compute dtype with --dtype {bf16,fp16,fp32}. CPU defaults to fp16 (PyTorch 2.x has CPU fp16 kernels); use --dtype fp32 if a fp16 op is unsupported or slower on your hardware.

4-bit / 8-bit quantization (MLX only): --quantize_bits {4,8} quantizes the T3 LLaMA backbone via mlx.nn.quantize (the S3Gen vocoder and voice encoder stay in --dtype). Tune the group size with the CHATTERBOX_FLASH_MLX_QUANT_GROUP env var (default 64). There is no CPU quantization path — PyTorch CPU has no native 4-bit kernels; use bf16/fp16 on CPU, or MLX quantization on Apple Silicon.

# Apple Silicon, 4-bit quantized backbone
python synthesize.py --audio_prompt reference.wav --text "..." \
    --backend mlx --quantize_bits 4

# Default — GPU + torch SDPA
python synthesize.py --audio_prompt reference.wav \
    --text "Sometimes it's better to just let things slide, you know?"

# GPU + FlashInfer (paged KV + CUDA graph)
python synthesize.py --audio_prompt reference.wav --text "..." --backend flashinfer

# CPU only
python synthesize.py --audio_prompt reference.wav --text "..." --backend cpu

# Apple Silicon native Metal
python synthesize.py --audio_prompt reference.wav --text "..." --backend mlx

# Multiple sentences; 8 rows per batched forward
python synthesize.py --audio_prompt reference.wav \
    --text "First sentence." "Second sentence." "Third sentence." \
    --batch_size 8

# From a file (one sentence per line, '#' lines and blanks ignored)
python synthesize.py --audio_prompt reference.wav --text_file sentences.txt

Defaults reproduce the paper's best decoding setup. CFG, when on (--cfg_scale > 0), is locked to the production combination — zero_text_batch

• zero_all null + null-text zeroed + null-speech duplicated + PMI combined via pmi_cfg ((1+w)·pmi_c − w·pmi_u) — no other CFG mode is exposed. Other defaults: OmniVoice r_n schedule (omnivoice_schedule_t_shift=0.5), position temperature T=5, precomputed unconditional block prior, S3Gen meanflow vocoding at 2 CFM steps. Override per run via --num_steps, --temperature, --time_shift_tau, --cfg_scale, --n_cfm_timesteps, --position_temperature.

Python API

import torchaudio as ta
from chatterbox_flash import ChatterboxFlashTTS

model = ChatterboxFlashTTS.from_pretrained("ResembleAI/chatterbox-flash", device="cuda")
wav = model.generate(
    "Sometimes it's better to just let things slide, you know?",
    audio_prompt_path="reference.wav",
)
ta.save("out.wav", wav, model.sr)

See examples/ for more demos.

Reproducing the paper benchmarks

# Downloads the released checkpoint from the Hugging Face Hub + OmniVoice's
# eval datasets/models, generates wavs, then scores SIM-o / WER / UTMOS using
# the same code OmniVoice publishes.
bash scripts/run_eval.sh

Defaults reproduce the Table 1 configuration from our paper (block size 16, max 10 steps per block, temperature 0.6, time-shift τ=0.1, CFG scale 1.0, zero-text-batch + pmi_cfg + zero_all null prefix).

Project layout

synthesize.py             ← single CLI entry point; one ref + many texts; --backend selects engine

chatterbox_flash/
├── model.py              ChatterboxFlashT3: chatterbox T3 + MASK token + block-diffusion generate()
├── tts.py                ChatterboxFlashTTS — user-facing pipeline (T3 + S3Gen + VE + tokenizer)
├── cfg_guidance.py       Classifier-free guidance helpers (zero-text-batch + PMI-side combination)
├── calibration.py        Prior-calibrated PMI scoring
├── engines/              Pluggable inference backends
│   ├── base.py             InferenceEngine protocol
│   ├── flashinfer.py       Paged KV + CUDA-graph FlashInfer engine (CUDA, preferred)
│   ├── torch_sdpa.py       Pure-PyTorch SDPA + DynamicCache fallback (CPU/CUDA/MPS)
│   ├── mlx.py              Apple-Silicon Metal engine via mlx + mlx-lm (experimental)
│   └── __init__.py         build_engine() — picks the backend, honours $CHATTERBOX_FLASH_ENGINE
├── text_norm/            English text normalization (numbers, abbreviations, dates, times, phones)
└── eval/                 OmniVoice JSONL generation + WER (Whisper, seedtts-eval style)

scripts/
└── run_eval.sh           Full evaluation pipeline (download → generate → score)

Citation

@article{chatterboxflash2026,
  title   = {Chatterbox-Flash: Prior-Calibrated Block Diffusion for Streaming Zero-Shot TTS},
  author  = {Resemble AI},
  year    = {2026},
}

License

MIT (see LICENSE).

The base architecture, S3Gen vocoder, voice encoder and tokenizer are provided by chatterbox-tts (also MIT-licensed).