Chatterbox-Flash Demo Samples
Zero-shot TTS — same quality as Chatterbox, decoded in parallel.
Chatterbox-Flash is a prior-calibrated block-diffusion zero-shot TTS model. It replaces the autoregressive T3 decoder of Chatterbox-TTS with a Fast-dLLM v2-style parallel masked decoder, while reusing the same S3Gen flow-matching vocoder and voice encoder unchanged. Instead of generating one speech token at a time, it decodes whole blocks of tokens in parallel — preserving Chatterbox's zero-shot voice fidelity while running ~2× faster than a vLLM-optimized Chatterbox on an H100.
Two inference-time techniques from our paper keep quality high at low step counts: prior-calibrated PMI scoring (suppresses the dominant-token bias of discrete speech codecs) and early decoding via a time-shifted quantile schedule (stops iterating once the calibrated confidence is high). The released package is inference-only and MIT-licensed.
Inference Speed
Measured on a single NVIDIA H100, the parallel block-diffusion decoder roughly halves the real-time factor (RTF) — even against a vLLM-optimized Chatterbox baseline:
| Model | RTF ↓ | TTFP ↓ |
|---|---|---|
| Chatterbox (vLLM-optimized) | 0.226 | 257 ms |
| Chatterbox-Flash | 0.107 | 118 ms |
RTF = generation time ÷ audio duration; TTFP = time-to-first-packet (both lower is better). Measured on a single NVIDIA H100; the Chatterbox baseline is served with vLLM optimizations.
Expressive Speech — Chatterbox vs Chatterbox-Flash
Same reference voice and same text for both models, so you can compare quality directly. Prompts and Chatterbox outputs are the originals from the Chatterbox demo.
Text: "My name is Maximus Decimus Meridius, commander of the Armies of the North, General of the Felix Legions and loyal servant to the true emperor, Marcus Aurelius. Father to a murdered son, husband to a murdered wife. And I will have my vengeance, in this life or the next."
| Prompt | Chatterbox | Chatterbox-Flash |
|---|---|---|
Text: "Introducing the next generation of refreshment. Duff Beer just got bolder, smoother, and brewed to perfection. Whether you're kicking back or barfing up, it's the taste that never quits. Crack open a classic. Duff is back, and better than ever."
| Prompt | Chatterbox | Chatterbox-Flash |
|---|---|---|
Text: "So I want you to get up now. I want all of you to get up out of your chairs. I want you to go to the window, open it, and stick your head out and yell 'I'M MAD AS HELL, AND I'M NOT GOING TO TAKE THIS ANYMORE!'"
| Prompt | Chatterbox | Chatterbox-Flash |
|---|---|---|
Text: "Every day I carry her name like a shield, and every night I wonder what I'm defending. Shar doesn't ask for love, only obedience, but sometimes I dream of light, and when I wake, I feel guilty for missing it."
| Prompt | Chatterbox | Chatterbox-Flash |
|---|---|---|
Text: "So let's make my mum's favourite. So three mars bars into the pan. NO don't touch it Rupert! Then we add the tuna and just stir for a bit, just let the chocolate and fish infuse. Then, a sprinkle of olive oil and for the finishing touch, some tomato ketchup. And there you have it, take that Gordon Ramsey!"
| Prompt | Chatterbox | Chatterbox-Flash |
|---|---|---|
Text: "The point is, ladies and gentlemen, that greed, for lack of a better word, is good. Greed is right. Greed works. Greed clarifies, cuts through, and captures the essence of the evolutionary spirit. Greed, in all of its forms, greed for life, for money, for love, knowledge, has marked the upward surge of mankind. And greed, you mark my words, will not only save Teldar Paper, but that other malfunctioning corporation called the USA."
| Prompt | Chatterbox | Chatterbox-Flash |
|---|---|---|
Code & weights: GitHub · Hugging Face — Made with ♥️ by Resemble AI