Hacker News

Mistral’s Voxtral Realtime 4B speech-to-text model represents a significant advancement in real-time audio processing, delivering cutting-edge features and performance. Built on a pure C implementation, this model ensures compatibility with zero external dependencies beyond the standard C library. Notably, it includes Metal GPU acceleration tailored for Apple Silicon, enabling high efficiency and speed.
Voxtral Realtime offers streaming output with incremental audio processing, which is supported by a memory-mapped BF16 weights system and live microphone input. A rolling KV cache allows for the handling of audio of unlimited length. This system efficiently transcribes with ~284ms encoder time for 3.6 seconds of audio and achieves ~23.5ms per decoder step on an M3 Max, operating 2.5x faster than real-time.
Additionally, Voxtral features a standalone C inference engine alongside a simple Python reference implementation. For developers interested in open-source solutions, Voxtral Realtime is provided under the Apache 2.0 license, contributing to the community with a high-performance, low-latency option.
Aside from real-time transcription, Mistral AI has introduced the Voxtral Transcribe 2 suite, which includes Voxtral Mini Transcribe V2 for batch processing. The models together support 13 languages, speaker diarization, and word-level timestamps, achieving state-of-the-art accuracy at a cost-effective rate of $0.003-0.006 per minute. These models outperform competitors like GPT-4o mini and Gemini 2.5 Flash in terms of price-performance ratio.
Voxtral’s comprehensive feature set includes context biasing for domain-specific vocabulary, support for audio files up to 3 hours in length, and compliance with GDPR and HIPAA standards. Mistral Studio’s new audio playground further enhances the user experience by enabling instant testing of audio input. These capabilities make Voxtral a leading choice in the rapidly evolving field of speech-to-text technology.

Collections

Mistral released Voxtral Realtime 4B, a speech-to-text model with pure C implementation, Metal GPU acceleration for Apple Silicon, and streaming output capabilities. It processes audio 2.5x faster than real-time with ~284ms encoder time for 3.6 seconds of audio on M3 Max. The Apache 2.0 licensed model features unlimited audio length support via rolling KV cache, zero external dependencies, and includes both C inference engine and Python reference implementation. Mistral also introduced Voxtral Transcribe 2 suite supporting 13 languages, speaker diarization, word-level timestamps, and GDPR/HIPAA compliance at $0.003-0.006 per minute, outperforming GPT-4o mini and Gemini 2.5 Flash on price-performance.

Voxtral: Advanced Real-Time Speech-to-Text Processing with Comprehensive Features