Summary
Voxtral TTS is an expressive multilingual text-to-speech model that generates natural speech from just 3 seconds of reference audio. The model uses a hybrid architecture combining autoregressive generation of semantic speech tokens with flow matching for acoustic tokens. Tokens are encoded and decoded by the Voxtral Codec, trained from scratch with a hybrid VQ-FSQ quantization scheme. In human evaluations conducted by native speakers, Voxtral TTS was preferred for naturalness and expressiveness in multilingual voice cloning, achieving a 68.4% win rate against ElevenLabs Flash v2.5. The paper was authored by Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg, Andy Lo, Chen-Yo Sun and colleagues, and released on arXiv (2603.25551) on March 26, 2026.
Paper Overview
Research area: Machine Learning
Authors: Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg, Andy Lo, Chen-Yo Sun, et al.
Published: 2026-03-26
arXiv: 2603.25551
Abstract
This paper introduces Voxtral TTS, an expressive multilingual text-to-speech model capable of generating natural speech using only 3 seconds of reference audio.
Architecture
Voxtral TTS adopts a hybrid architecture that combines:
- Autoregressive generation of semantic speech tokens
- Flow matching for acoustic tokens
These tokens are encoded and decoded using the Voxtral Codec, which is trained from scratch and employs a hybrid VQ-FSQ quantization scheme.
Evaluation Results
In human evaluations conducted by native speakers, Voxtral TTS was preferred for its naturalness and expressiveness in multilingual voice cloning, achieving a 68.4% win rate relative to ElevenLabs Flash v2.5.
---
*Auto-collected on 2026-03-29*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177169393