Voxtral Realtime

TL;DR

Voxtral Realtime is a streaming speech recognition model trained end-to-end for sub-second latency, achieving offline-quality transcription across 13 languages without chunking or sliding windows.

Contribution

It introduces a novel end-to-end streaming ASR architecture with explicit audio-text alignment and scalable multilingual pretraining.

Findings

Achieves offline-quality transcription at 480ms delay.

Performs on par with Whisper across 13 languages.

Released under Apache 2.0 license.

Abstract

We introduce Voxtral Realtime, a natively streaming automatic speech recognition model that matches offline transcription quality at sub-second latency. Unlike approaches that adapt offline models through chunking or sliding windows, Voxtral Realtime is trained end-to-end for streaming, with explicit alignment between audio and text streams. Our architecture builds on the Delayed Streams Modeling framework, introducing a new causal audio encoder and Ada RMS-Norm for improved delay conditioning. We scale pretraining to a large-scale dataset spanning 13 languages. At a delay of 480ms, Voxtral Realtime achieves performance on par with Whisper, the most widely deployed offline transcription system. We release the model weights under the Apache 2.0 license.