Rednote drops dots.tts, an open-source 2B speech model

dots.tts is an open-source 2B-parameter text-to-speech system from Rednote designed for zero-shot voice cloning and precision speech editing.

Hand it a short reference audio clip, and it can mimic the speaker's vocal traits to read out any text with impressive fidelity, or edit words, emotions, and cadence directly inside an existing recording.

A fully continuous setup with zero discrete tokens

Most modern autoregressive speech models convert audio into discrete acoustic tokens, essentially treating speech generation like text prediction. dots.tts skips discrete tokens entirely.

Its architecture connects a semantic encoder, an LLM backbone, and an autoregressive flow-matching acoustic head operating over a 48 kHz AudioVAE. The entire pipeline remains continuous.

By avoiding discrete tokenization, the model skips lossy quantization, preserving fine vocal timbre and natural acoustic nuances that discrete models often lose.

Benchmarks and real-world throughput

On the standard Seed-TTS-Eval benchmark, dots.tts achieved state-of-the-art results among open-source models, recording word error rates of 0.94% on Chinese, 1.30% on English, and 6.60% on hard Chinese sets. Speaker similarity scores came in at 81.0, 77.1, and 79.5 respectively.

It also hit an average speaker similarity of 83.9 across MiniMax's 24-language multilingual evaluation. The team also released dots.tts.edit, an extension focused on instruction-controlled editing for words, prosody, and emotion.

On inference speed, the optimized dots.tts-mf checkpoint delivers a real-time factor p50 of 0.13 to 0.15, with first-chunk latency dropping to 68 ms for pure text and 204 ms for voice cloning. Integrated with SGLang Omni, a single H100 GPU running concurrency 16 pushes throughput up to 19.86 seconds of audio per second.

Getting started

The entire project is licensed under Apache-2.0, with inference code, fine-tuning scripts, and pretrained checkpoints publicly available. It requires Python 3.10 to 3.12.

You can grab it via pip install dots.tts or clone the repository directly from GitHub. The team provides base checkpoints, SOAR weights, and distilled MeanFlow variants.

The command-line interface supports continuation cloning with reference audio and transcript, standalone audio-only cloning, and sampling for single-speaker fine-tuned models. For high-concurrency production setups, pairing it with SGLang Omni is the recommended route.

Why it matters

If you build voice agents, podcast localization pipelines, or in-game dialog systems, dots.tts offers a rare mix of open weights, low latency, and high naturalness under a permissive Apache-2.0 license.

Most high-performing voice cloning tools remain locked behind proprietary APIs. Releasing a 2B fully continuous model that matches commercial-grade error rates and includes speech editing tools gives self-hosted voice infrastructure a serious upgrade.