Mistral launches affordable speech transcription model that runs on smartphones

Paris-based startup Mistral AI has unveiled two new Voxtral Transcribe 2 speech recognition models that run locally on users’ devices. They process audio up to ten times faster than competitors while costing five times less.

• The models handle audio directly on the device without sending data to servers, a crucial feature for industries like healthcare, finance, and government
• API transcription costs $0.003 per minute for batch processing, and $0.006 per minute for real-time use
• Real-time latency can be tuned down to 200 milliseconds, compared to Google’s two-second delay

Two models optimized for different tasks

Mistral split its tech into two distinct offerings. Voxtral Mini Transcribe V2 handles batch processing of recorded files and supports 13 languages including English, Mandarin, Japanese, Arabic, Hindi, and several European tongues. The company claims it delivers among the lowest word error rates in the transcription market. API access costs just $0.003 per minute-five times cheaper than major competitors.

Voxtral Realtime is designed for live audio with latency as low as 200 milliseconds. This model is open source under the Apache 2.0 license-developers can download the weights from Hugging Face, modify, and deploy them without paying Mistral licensing fees. For those who prefer a managed solution, the API costs $0.006 per minute.

“The open-source community is incredibly inventive when it comes to applications. We’re excited to see what they’ll build,” said Pierre Stock, Mistral’s VP of Scientific Operations, in an interview with VentureBeat.

On-device processing tackles privacy concerns

The push to make models compact enough to run locally aligns with where the enterprise market is heading. As companies increasingly use AI for sensitive workflows like transcribing medical consultations, financial advisor calls, and legal testimonies, controlling data flow becomes critical.

Stock pointed out a common issue with current audio note-taking apps: they indiscriminately pick up background noise. That can mean accidentally transcribing songs playing nearby, other people’s conversations, or hallucinating text because of ambient sounds. Mistral invested heavily in curating training data and refining the model architecture to address these problems.

Mistral also added enterprise-friendly features that some US rivals have adopted more slowly. Contextual biasing lets customers upload specialized terminology lists-medical jargon, product names, industry acronyms-and the model prioritizes these when transcribing ambiguous audio. Unlike fine-tuning, which requires retraining the model, contextual biasing works through a simple API parameter.

“You just provide a plain text list, and the model automatically shifts transcription toward those abbreviations or unusual words. It works without examples, no retraining, no complicated tweaks needed,” Stock explained.

Weighing just 4 billion parameters, the model is small enough to run on laptops, smartphones, or smartwatches. This lets users process voice and transcription locally, without sending data to remote servers-a key advantage for regulated sectors like healthcare, finance, and defense.

From factory floors to call centers

Stock described two main use cases. The first covers industrial audits: technicians walk around factories, inspecting heavy machinery and shouting notes amid loud ambient noise. The challenge is capturing specialized technical vocabulary that only these experts know how to spell correctly, producing timestamped notes with speaker identification and high noise resilience.

The second focuses on customer support operations. When a customer calls, Voxtral Realtime transcribes the conversation instantly and feeds the text into backend systems that retrieve relevant account info before the caller finishes explaining the issue.

“The status pops up on the agent’s screen before the customer even finishes their sentence or stops complaining. That way, you can interact immediately and say: ‘Okay, I see your status. Let me fix the address and resend the package,'” Stock said. He estimated this could shrink typical support interactions to just two exchanges: the customer explains the problem, and the agent resolves it immediately.

Real-time translation could arrive by the end of 2026

While the focus is on transcription now, Stock clarified that Mistral views these models as foundational technology for a more ambitious goal: natural, real-time speech-to-speech translation.

“The ultimate target application could be live translation. I speak French, you speak English. Minimizing delay is critical, otherwise you lose empathy. Your facial expressions won’t sync with what you said a second ago,” he said.

This sets Mistral in direct competition with Apple and Google, who are also chasing this challenge. Google’s latest translation model has a two-second latency-ten times slower than Mistral’s claimed 200 milliseconds with Voxtral Realtime.

Mistral positions itself as a privacy-first corporate alternative

Mistral occupies a unique spot in the AI landscape. Founded in 2023 by ex-Meta and Google DeepMind engineers, the startup has raised over $2 billion and is currently valued at about $13.6 billion. Still, it operates with a fraction of the compute power available to American hyperscalers, building its strategy around efficiency rather than brute force.

“Our models are enterprise-grade, industry-leading, efficient-especially cost-wise-and can be deployed at the edge, unlocking privacy, control, and transparency,” Stock said.

This privacy-centric approach resonates strongly with European clients wary of dependence on US tech. In January, France’s Ministry of Armed Forces signed a framework deal giving the country’s military access to Mistral’s AI models-a partnership clearly requiring deployment on France-controlled infrastructure.

Data privacy remains one of the biggest hurdles for corporate adoption of voice AI. For companies in sensitive sectors-finance, manufacturing, healthcare, insurance-sending audio data to external cloud servers is often unacceptable. Information must stay on-device or within the company’s own infrastructure.

Competing with OpenAI, Google, and China’s growing presence

The transcription market is fiercely competitive. OpenAI’s Whisper model has become an industry standard, available through API and as open-source weights. Google, Amazon, and Microsoft offer enterprise speech services, while niche players like Assembly AI and Deepgram have built solid businesses serving developers needing scalable transcription.

Mistral claims its new models outperform all rivals on accuracy tests while cutting costs. “We beat them on benchmarks,” Stock said. Independent verification will take time, but the company points to FLEURS-a widely used multilingual speech benchmark-where Voxtral models achieve word error rates on par with or better than those from OpenAI and Google alternatives.

Perhaps more importantly, Mistral’s CEO Arthur Mensch warned that US AI companies face pressure from an unexpected quarter. Speaking at the World Economic Forum in Davos last month, Mensch dismissed the notion that Chinese AI lags behind the West as a “myth.”

“The open-source tech capabilities coming out of China probably make US CEOs nervous,” he said.

French startup bets on trust

Stock predicted that 2026 will be the “year of note-taking” – when AI transcription becomes reliable enough that users fully trust it.

“You need to trust the model, and the model basically can’t make a single mistake, or you lose trust in the product and stop using it. The bar is super, super high,” he said.

Whether Mistral has crossed that threshold remains to be seen. Enterprise clients will be the ultimate judges, and they tend to move slowly, vetting claims before allocating budgets and workflows to new tech. The audio playground in Mistral Studio, where developers can test Voxtral Transcribe 2 with their own files, launched today.

But Stock’s broader point is worth noting. In a market where American giants throw billions at ever-larger models, Mistral is taking a different route: in the AI era, smaller and more local can beat bigger and more centralized. For executives worried about data sovereignty, regulatory compliance, and vendor lock-in, this pitch might be more persuasive than any benchmark score.

The race to dominate corporate voice AI is no longer just about who builds the most powerful model – it’s about who builds the model you’re willing to let listen.

* Meta is recognized as an extremist organization and banned in Russia.