Mistral just threw down the gauntlet in the speech AI world with Voxtral—a bold open-source alternative to OpenAI Whisper and ElevenLabs. With high benchmarks, multilingual support, and ultra-low pricing, it’s making waves in the AI developer community.
Key Takeaways:
- Open-source speech model: Voxtral rivals Whisper & ElevenLabs
- Ultra-low cost: Just $0.001/minute for transcription
- Multilingual & context-rich: Up to 32K token context
- Enterprise-ready: On-premise, emotion detection, speaker ID
- Human-like voice interface push: Mistral hiring for expansion
In a strategic move shaking up the speech AI landscape, French AI startup Mistral has unveiled Voxtral, a new family of open-source speech understanding models aimed directly at the strongholds of OpenAI Whisper and ElevenLabs Scribe.
Available in two sizes—a 24 billion parameter model for enterprise-scale operations and a lighter 3B “Mini” for edge or local use—Voxtral is free to use under the Apache 2.0 license and accessible via Hugging Face or Mistral’s API.
But what really turns heads? Its ultra-affordable transcription endpoint at just $0.001 per minute. That’s not just competitive—it’s disruptive.
Voxtral is built to handle long-form audio with a whopping 32,000-token context window, enabling complex capabilities like direct question answering, summarization, and even triggering actions from voice commands, all in real-time and without chaining models.
And this isn’t just hype. Benchmark tests released by Mistral show Voxtral outperforming Whisper Large V3, GPT-4o Mini Transcribe, and Gemini 2.5 Flash on transcription quality and multilingual accuracy, especially in European languages.
The backbone? Voxtral retains the text power of Mistral Small 3.1, allowing seamless switching between voice and language tasks. For enterprise customers, Mistral goes further—offering on-prem deployment, domain-specific fine-tuning, and advanced features like speaker ID, emotion detection, and diarization.
Developers can already test the models in Le Chat’s voice mode or through Mistral’s API, with a full demo planned on August 6 via webinar with Inworld AI.
This comes hot on the heels of Mistral’s Magistral model, focused on complex reasoning in sectors like healthcare and finance. Together, they show Mistral isn’t just participating in the AI race—it’s aiming to lead.
As the startup scales up its audio team to build near-human-like voice interfaces, one thing’s clear: Open-source AI is back in the spotlight—and Mistral wants the mic.