hugging-apps/dots-tts-demo

dots.tts — zero-shot voice cloning

Demo of [dots.tts](https://github.com/rednote-hilab/dots.tts) (paper 2606.07080), a 2B-parameter fully continuous end-to-end autoregressive TTS system: a semantic encoder, an LLM and an autoregressive flow-matching acoustic head over a 48 kHz AudioVAE.

This Space runs the MeanFlow-distilled 2-step checkpoint `dots-studio/dots.tts-mf-2steps`. Its sampling contract (Euler ODE, 2 steps, guidance 0.0) is locked in the checkpoint config, which is why those knobs are not exposed.

How to use

1. 1.Upload ~10 s of prompt audio (the voice to clone, ≤ 30 s).
2. 2.Enter its exact transcript — mismatches degrade cloning stability.
3. 3.Type the text the voice should speak and hit Generate.

Different seeds give different rhythm and intonation; try a few.

Example audio

The bundled example clips come from the seed-tts-eval dataset (CC-BY-4.0), the same benchmark the dots.tts authors evaluate on.