New models challenge Whisper and Gemini in multilingual voice tasks
French AI innovator Mistral has entered the speech tech arena with the release of Voxtral, its first family of open-source models for audio-based tasks. Designed for transcription, speech translation, summarization, and multilingual audio understanding, Voxtral positions Mistral as a serious new contender in the voice AI space — challenging players like OpenAI, Google, and ElevenLabs.
Three Models, Multiple Use Cases
The Voxtral family includes Voxtral Small, a 24.3B parameter model built for production-scale workloads, and Voxtral Mini, a 4.7B variant optimized for local and edge use. For lightweight needs, Voxtral Mini Transcribe offers a fast, cost-efficient alternative tailored exclusively for transcription.
What sets Voxtral apart is its capacity: Voxtral Small and Mini can handle audio files up to 40 minutes long, while Voxtral Mini Transcribe supports files up to 30 minutes. These extended limits open up new possibilities in areas like multilingual meeting transcription, voice-driven chat interfaces, and long-form audio summarization — tasks where most other models, open or closed, fall short.
Outperforming the Giants
In internal testing, Mistral reported that Voxtral Small and Voxtral Mini Transcribe delivered state-of-the-art transcription results, outperforming OpenAI’s Whisper, GPT-4o Mini Transcribe, and Google’s Gemini 2.5 Flash, and holding its own against ElevenLabs Scribe. In speech translation tasks, Voxtral Small beat both Gemini 2.5 Flash and GPT-4o Mini Audio across several language pairs, including English↔French, Spanish↔English, and German↔English.
Open Source, Benchmarked, and Ready to Use
Mistral has made the model weights for both Voxtral Small and Voxtral Mini available on Hugging Face, along with a free API and a live demo via its chatbot Le Chat.
In addition, the company introduced three new speech understanding benchmarks derived from well-known text evaluation datasets. These synthetic speech benchmarks, released under a permissive license, aim to promote more consistent and transparent evaluations across the industry. “We encourage their adoption as standard benchmarks for speech understanding,” the team stated.
A Strategic Expansion
Known for its rapid rise in the open-source LLM space, Mistral’s expansion into voice signals a strategic push to broaden its portfolio and compete in multimodal AI. With Voxtral, the company is not just adding speech capabilities — it’s setting a new bar for open, multilingual, and scalable audio AI.
