Voxtral transcribes

lastmod 2026-02-05created 2026-02-051 edits

Speech-to-text model

Voxtral Transcribe 2

Next-generation speech-to-text models with ultra-low-latency real-time speech recognition and high-precision speaker diarization

Consists of two models: Voxtral Mini Transcribe V2 for batch processing, and Voxtral Realtime, built for real-time applications
The Realtime model processes speech in a streaming manner with under 200ms latency, and is released as Apache 2.0 open weights
Mini Transcribe V2 supports 13 languages including Korean and provides enterprise features such as word-level timestamps, context biasing, and speaker diarization
Both models support GDPR- and HIPAA-compliant deployment, greatly improving the accuracy, speed, and cost efficiency of voice-based applications

Voxtral Transcribe 2 consists of two models featuring state-of-the-art speech recognition quality, precise speaker diarization , and ultra-low-latency processing
- Voxtral Mini Transcribe V2: for batch transcription
- Voxtral Realtime: for real-time applications
The Realtime model is released under the Apache 2.0 license and can be deployed even in edge environments
Transcription can be tested immediately through the audio playground in Mistral Studio

Voxtral Mini Transcribe V2: supports 13 languages, provides speaker diarization, context biasing, and word-level timestamps
Voxtral Realtime: real-time transcription with under 200ms latency, suited to voice agents and real-time applications
Efficiency: top-tier accuracy at the lowest price in the industry
Open weights: the Realtime model is released under Apache 2.0, enabling privacy-focused deployment

A model designed for latency-critical applications; instead of processing audio in chunks, it performs real-time transcription with a streaming architecture
Latency can be configured below 200ms; at 2.4s latency it matches the accuracy of the batch model, and at 480ms latency it stays within a 1-2% error rate
Supports 13 languages (English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, Dutch)
At 4B parameters it runs efficiently even on edge devices, ensuring security and privacy
Model weights are published on the Hugging Face Hub

Transcription and diarization quality greatly improved across languages and domains
About 4% word error rate on the FLEURS benchmark, offering top-tier price-performance at $0.003/minute
More accurate than GPT-4o mini Transcribe, Gemini 2.5 Flash, Assembly Universal, and Deepgram Nova, and 3x faster than ElevenLabs Scribe v2 at about 1/5 the cost

Speaker diarization : distinguishes speakers and marks start/end times, suited to meetings, interviews, and multi-party calls
Context biasing : up to 100 words/phrases can be specified, improving recognition of proper nouns and technical terms (optimized for English, experimental for other languages)
Word-level timestamps: used for subtitle generation, audio search, and content alignment
Expanded language support: supports 13 languages, outperforming competing models even in non-English languages
Noise robustness: maintains accuracy even in noisy environments such as factories and call centers
Long-form audio processing: recordings of up to 3 hours can be processed in a single request

Voxtral Transcribe 2 can be tested directly in Mistral Studio
Upload up to 10 audio files; supports settings for speaker diarization, timestamp granularity, and context biasing
Supported formats: .mp3, .wav, .m4a, .flac, .ogg, up to 1GB per file

Meeting intelligence: multilingual meeting transcription and speaker separation enable analysis of large-scale meeting data
Voice agents and virtual assistants: natural conversational interfaces with under 200ms latency
Contact center automation: real-time call transcription supports sentiment analysis, response suggestions, and automatic CRM entry
Media and broadcasting: real-time multilingual subtitle generation, enhanced recognition of proper nouns and technical terms
Compliance and documentation: audit trails based on per-speaker timestamps
Both models support GDPR- and HIPAA-compliant deployment and can be operated securely in on-premises or private cloud environments

Voxtral Mini Transcribe V2: $0.003/minute via the API, available in Mistral Studio or Le Chat
Voxtral Realtime: $0.006/minute via the API, open weights available on Hugging Face
More information is available in Mistral's audio and transcription documentation