Chatterbox-Flash

Chatterbox-Flash

A zero-shot English text-to-speech model that copies a voice from a short reference clip and streams audio block by block with low latency. Suits developers building voice assistants and narration tools.

Made by ResembleAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Chatterbox-Flash is a text-to-speech model from Resemble AI that works zero-shot, meaning you give it a short reference recording and it speaks new text in that voice. It decodes several speech tokens in parallel inside each block and streams audio block by block. The card reports quality close to slower autoregressive systems at much lower latency.

Who it suits: Developers building voice assistants, narration tools or live speech features in English, where audio should start fast. Use only voices you have the right to use, since it copies the voice from the reference clip.

Industry | Who uses it | How they use it | Example
Software | Voice app developers | Stream spoken replies as they are generated | Speak an assistant answer quickly
Media | Narration and audio teams | Turn scripts into speech in a chosen voice | Read a script in your own recorded voice
Accessibility | Product teams | Add spoken output to text interfaces | Read on-screen text aloud
Research | Speech researchers | Study block-diffusion text-to-speech | Compare decoding settings for speed

Anyone who needs languages other than English can skip it, since the model page lists English only.

Where to use it:

• Python scripts that use the chatterbox-flash package, on a CUDA GPU as in the card's example
• Apple Silicon Macs, through MLX, with 4-bit quantization for extra speed
• Windows, macOS or Linux, after downloading the files with Hugging Face's hub tool

Use case | Who | What they get back
Streaming speech | Voice app developers | Audio packets as each block is finished
Voice cloning from a clip | Audio creators | New speech in the style of a reference recording you have rights to
Local Mac synthesis | Mac users | Speech faster than real time on an M4, per the card

How to set it up:

• Python with pip
• A CUDA GPU for the card's example, or an Apple Silicon Mac for the MLX route
• A short reference audio file of a voice you may use

1. Install the Hugging Face hub package and download the files, if you want them on disk.
2. Install the chatterbox-flash package with pip.
3. Run the snippet below with your own reference file.

pip install -U huggingface_hub
hf download ResembleAI/chatterbox-flash

pip install chatterbox-flash

How to use it: Load the model onto the GPU, then call generate with your text and the path to a reference recording. The snippet saves the result as an audio file.

import torchaudio as ta
from chatterbox_flash import ChatterboxFlashTTS
tts = ChatterboxFlashTTS.from_pretrained("ResembleAI/chatterbox-flash", device="cuda")
wav = tts.generate(
    "Hello, world.",
    audio_prompt_path="reference.wav",
)
ta.save("out.wav", wav.unsqueeze(0).cpu(), tts.sr)

Tool | What it does
ChatterboxFlashTTS.from_pretrained | Loads the model onto the device you choose
generate | Speaks the text, using the reference clip as the voice prompt
ta.save | Writes the generated audio to a file

File | What it is
t3_flash.safetensors | The block-diffusion decoder, based on a 520M Llama with one extra mask token
s3gen.safetensors | The flow-matching vocoder
ve.safetensors | The voice encoder, taken from the original Chatterbox
tokenizer.json | The English tokenizer, also from the original Chatterbox

Example: You run the snippet with the text "Hello, world." and a reference recording named reference.wav. It writes out.wav, a spoken version of that text in the style of the reference voice.

For newcomers and experts: If you are new, run the snippet once with a short clear reference clip and listen to out.wav. Then change the text and compare how closely the voice matches.

Experts get a block-diffusion decoder that unmasks multiple speech tokens at once in each block, with an early-decoding schedule that cuts average steps per block by about 20 percent. Defaults are block size 16, up to 10 denoising steps per block, temperature 0.2, and a FlashInfer paged KV cache with CUDA graph capture.

Title: Real-time factor, lower is faster
Default setting: 0.107 RTF
Fastest setting: 0.076 RTF
Mac M4 with MLX: 0.778 RTF
Mac M4 with MLX 4-bit: 0.665 RTF

You give text and a reference clip
The decoder fills each block in parallel
The vocoder turns tokens into audio
Audio streams out block by block

Good to know:

• Speed: about 9 times real time at the default setting and about 13 times at the fastest, per the card
• Latency: time to first audio packet is 103 to 119 ms across the tested settings
• Language: English only
• License: MIT
• Cost to run: none named, no API key or account, though the example needs a CUDA GPU
• Data it reaches: it downloads weights from Hugging Face and reads the reference audio and text you give it
• Not for: using a voice you have no right to use, and it does not cover languages other than English

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: chatterbox-flash
• Downloads in the last 30 days: 0
• Likes: 33
• License: mit
• Languages: en
• Quantized versions: 1
• Fine-tunes: 2
• Tags: chatterbox-flash, text-to-speech, zero-shot-tts, block-diffusion, speech-synthesis

Links

• Model: https://huggingface.co/ResembleAI/chatterbox-flash
• Files: https://huggingface.co/ResembleAI/chatterbox-flash/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ResembleAI/chatterbox-flash

Then follow the usage in its model card. Nothing in this kit runs by itself.