Mistral drops new speech-to-text AI models

2025-12-15 · Source: aibusiness · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, quick

Summary

Mistral, a French AI startup, has launched two new speech-to-text models, Voxtral Mini Transcribe V2 and Voxtral Realtime, under the Voxtral Transcribe 2 umbrella. These models aim to establish new benchmarks for speed, privacy, and affordability in transcription. Voxtral Realtime is designed for live audio processing, featuring a novel streaming architecture that delivers transcriptions with configurable delays as low as 200 milliseconds and supports 13 languages. It has four billion parameters, enabling on-device deployment for enhanced privacy and security, and is available under an Apache 2.0 license or via API at $0.006 per minute. Voxtral Mini Transcribe 2 focuses on batch transcription of pre-recorded audio, offering speaker diarization, context biasing, and word-level timestamps. It supports the same 13 languages and boasts a 4% word error rate on the FLEURS benchmark, priced at $0.003 per minute.

Key takeaway

For AI Architects and Machine Learning Engineers evaluating speech-to-text solutions, Mistral's Voxtral models offer compelling options. Consider Voxtral Realtime for applications requiring ultra-low latency live transcription and on-device privacy, or Voxtral Mini Transcribe 2 for cost-effective batch processing with advanced diarization. Evaluate their performance and pricing against your specific project requirements, especially for multilingual and privacy-sensitive deployments.

Key insights

Mistral's new speech-to-text models prioritize speed, privacy, and affordability for diverse enterprise applications.

Principles

• On-device processing enhances privacy.
• Streaming architecture improves real-time transcription.
• Multilingual support expands application scope.

Method

Voxtral Realtime uses a novel streaming architecture for ultra-low latency live audio processing. Voxtral Mini Transcribe 2 handles batch processing with features like speaker diarization and context biasing.

In practice

• Deploy Realtime for voice agents or subtitling.
• Use Mini Transcribe 2 for compliance documentation.
• Leverage on-device models for sensitive data.

Topics

• Speech-to-text
• Mistral AI
• Real-time Transcription
• On-device AI
• Speaker Diarization

Best for: AI Architect, Machine Learning Engineer, NLP Engineer, AI Engineer, AI Product Manager, Tech Journalist

Related on AIssential

• Speech intelligence for enterprises - Mistral AI — Mistral AI offers open, customizable speech intelligence models for enterprises, enabling human-like voice interactions and accurate transcription.
• Voxtral transcribes at the speed of sound — Mistral's Voxtral Transcribe 2 offers both open-weights and API-based speech-to-text models with real-time and diarization features.
• Mistral AI Launches Text-to-Speech Model — Mistral AI's Voxtral TTS offers open-weight, multilingual, adaptable text-to-speech for enterprise and edge devices.
• How Mistral’s Voxtral TTS is Redefining Multilingual Voice Cloning with a Hybrid Autoregressive and Flow-Matching Architecture — Voxtral TTS uses a hybrid autoregressive and flow-matching architecture for high-fidelity multilingual voice cloning.
• Voxtral Realtime — Voxtral Realtime delivers offline-quality ASR with sub-second latency via a natively streaming, end-to-end trained architecture.
• SoundWise Launches Free Forever AI Audio and Video Transcription Tool for Unlimited Speech to Text Conversion — SoundWise offers a dual-engine AI transcription service, providing free, private, local processing alongside a faster, professional cloud-based option.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by aibusiness.