chatterbox FLASH • resemble TTS research

A new architecture for production-grade speech

Get started
2x
Faster than AR baseline on vLLM
dLLM
Diffusion-LLM architecture
MIT
License
Trusted by
the problem

Autoregressive TTS hits a throughput ceiling. Flash breaks through it.

Autoregressive models generate speech token by token — high quality, but fundamentally sequential. More output requires more hardware. Diffusion-LLM architectures parallelize generation, but most never reach production. Flash does.
WHAT IT DOES

Higher throughput. Same provenance guarantees.

More audio per second. Every clip still watermarked, traceable, and yours.

Diffusion-LLM architecture
Parallel generation replaces the autoregressive token-by-token loop — higher throughput on the same hardware without sacrificing output quality.
2× faster on vLLM
2× throughput over the autoregressive Chatterbox baseline on vLLM. More audio, same hardware budget.
Prior-subtraction technique
Improves generation quality within the diffusion framework. We believe approach generalizes to any dLLM TTS architecture.
Production-deployed
One of the first dLLM TTS models in production — benchmarked against real workloads, not just research benchmarks.
PerTh watermarking
Every clip is watermarked at generation — imperceptible to listeners, persistent through re-encoding, verifiable on demand.
MIT licensed
Run it on your own infrastructure. Audit the architecture. No vendor dependency.
HOW THE MODEL WORKS

Chatterbox, rebuilt on a diffusion-LLM backbone.

Autoregressive TTS generates speech one token at a time. Quality is high, but throughput is bounded by the sequential process — scaling output means scaling compute proportionally.
Flash replaces the autoregressive loop with a diffusion-LLM backbone. Generation is parallelized, which is why Flash runs 2× faster than the AR baseline on the same vLLM infrastructure.
The prior-subtraction technique addresses a quality problem that appears in diffusion-based speech generation. We believe it is architecture-agnostic and will publish the approach.
deployment

Run on the infrastructure you already use.

High-throughput serving

2–3× higher throughput than conventional autoregressive LLM-based TTS engines with vLLM.

VLLM NATIVE
Cloud API

Resemble managed platform. RESTful API with streaming. Business plan or higher for Voice Cloning API.
‍

MANAGED
Self-hosted

MIT licensed. Deploy on your own infrastructure. Full control over data and compute.

MIT • SELF-HOST
Frequently asked questions
What is Chatterbox Flash?
Chatterbox Flash is Chatterbox rebuilt on a diffusion-LLM architecture. It delivers 2× the throughput of the autoregressive Chatterbox baseline on vLLM, making it suited for high-volume production workloads where inference speed is the primary constraint.
What is a diffusion-LLM architecture and why does it matter for TTS?
Autoregressive TTS models generate speech one token at a time — sequentially, which bounds throughput. A diffusion-LLM backbone parallelizes generation, allowing more audio to be produced in the same time on the same hardware. Flash is one of the first production TTS models to use this architecture.
What is the prior-subtraction technique?
Prior subtraction is a novel technique Resemble developed to improve generation quality within the diffusion framework. It addresses a fidelity problem that appears in diffusion-based speech generation. Resemble believes the approach is architecture-agnostic and applicable to any dLLM TTS system. A research paper is in preparation.
How does Flash compare to the standard Chatterbox model?
Flash delivers 2× the throughput of the autoregressive Chatterbox baseline on vLLM. The quality bar is production-grade — Flash is deployed and benchmarked against real workloads, not just research benchmarks. For teams already running Chatterbox on vLLM, Flash is a drop-in upgrade for throughput.
Does Flash include PerTh watermarking?
Yes. Every clip Flash generates carries a PerTh watermark — imperceptible to listeners, persistent through re-encoding and format conversion, and verifiable on demand to confirm the audio's origin.
Can I self-host Flash?
Yes. Flash is MIT-licensed. Deploy on your own infrastructure with full control over data and compute. Access is also available through the Resemble managed platform via the cloud API.
Get complete generative AI security
Join thousands of developers and enterprises securing with Resemble AI
Contact us