Speed of sound? Sounds good.
French firm Mistral AI has unleashed Voxtral Transcribe 2, a new family of speech recognition models, that transcribes “at the speed of sound.”
Voxtral Transcribe 2 consists of two speech-to-text models with transcription quality, diarization, and ultra-low latency. The family includes Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live applications.
According to Mistral, Voxtral Realtime uses “a novel streaming architecture that transcribes audio as it arrives,” rather than adapting offline models. This enables latency “configurable down to sub-200ms,” a threshold that is critical for voice assistants, live captioning, and conversational AI.
Product strategy
The company also revealed that Voxtral Realtime is released as open weights under the Apache 2.0 license, allowing organizations to deploy it on their own infrastructure, including edge devices. This has significant implications for privacy-sensitive industries such as healthcare, finance, and government, where sending audio data to third-party clouds is often restricted.
This emphasis on open-source and integration is not incidental. As companies increasingly worry about vendor lock-in and data sovereignty, Mistral positions itself as an alternative to closed, US-based AI platforms.
Pricing is addressed candidly. It’s usage-based, starting around €5K/month ($5,896). This signals that Mistral is targeting mid-to-large organizations rather than individual developers, while still framing its services as competitive on cost.
Performance
Mistral claims Voxtral Mini Transcribe V2 achieves “approximately 4% word error rate on FLEURS” at a price of “$0.003/min,” which it describes as “the best price-performance of any transcription API.”
The company says the model outperforms offerings from GPT-4o mini Transcribe, Gemini 2.5 Flash, Assembly Universal, and Deepgram Nova on accuracy, while processing audio “approximately 3x faster than ElevenLabs’ Scribe v2” at “one-fifth the cost.”
If independently validated, these claims could disrupt a market where speech-to-text pricing has remained relatively high, particularly for multilingual and diarized transcription. Lower costs make it economically viable to transcribe large volumes of meetings, calls, and media archives that were previously too expensive to process.
Enterprising ideas
Beyond raw transcription, Voxtral Mini Transcribe V2 introduces features aimed squarely at enterprise use. These include speaker diarization with precise timestamps, context biasing to handle domain-specific vocabulary, word-level timestamps, and improved robustness in noisy environments such as “factory floors” and “busy call centers.”
The model supports recordings up to three hours in a single request and operates across 13 languages, including English, Chinese, Hindi, Arabic, and several European and Asian languages. Mistral notes that “non-English performance significantly outpaces competitors,” addressing a long-standing weakness in speech AI dominated by English-centric training data.
Mistral frames Voxtral as a foundational layer for multiple industries. Media and broadcast organizations can generate live multilingual subtitles, while regulated sectors can rely on diarization and timestamps for compliance and audit trails. Both Voxtral models support GDPR- and HIPAA-compliant deployments through on-premise or private cloud setups.
Where speech AI is headed
The Voxtral launch illustrates how speech AI is moving from novelty to infrastructure. The combination of open weights, aggressive pricing, and real-time performance suggests that competition is shifting away from who has the largest model toward who can deliver practical, deployable systems.
The success of Voxtral Transcribe 2 may hinge less on technical benchmarks and more on whether it delivers the cost savings and efficiency gains promised.
In December, Mistral had the wind in its sails with the launch of the Mistral 3 model family.