Voxtral Transcribe 2 Review: Mistral's Sub-200ms Speech-to-Text (2026)
Released February 4, 2026 · Reviewed June 14, 2026
At a Glance
On February 4, 2026, Mistral AI released Voxtral Transcribe 2 — a two-model speech-to-text family that enters the market with the lowest word error rate and price-per-minute of any mainstream transcription API. The family includes Voxtral Mini Transcribe V2 for batch transcription ($0.003/min with speaker diarization) and Voxtral Realtime for live applications (sub-200ms latency, $0.006/min, Apache 2.0 open weights).
Mistral's entry into STT follows Voxtral TTS (March 2026) and completes a full audio stack alongside their LLM and vision offerings. Voxtral Transcribe 2 is immediately available via API and testable in the new Mistral Studio audio playground — no waitlist, no enterprise contract.
Need AI voices, not just transcription? ElevenLabs turns text into realistic speech — 10,000 free characters/month.
What's New vs. Prior Voxtral
Sub-200ms latency via a novel streaming architecture
Voxtral Realtime doesn't work by adapting an offline batch model with audio chunking — a common approach that introduces irreducible latency as the model waits for each chunk boundary. Instead, Mistral built a purpose-designed streaming architecture that transcribes audio as it arrives. At 480ms delay, Voxtral Realtime stays within 1–2% word error rate of the batch model. At 2.4 seconds delay (ideal for subtitling), it matches Voxtral Mini Transcribe V2's accuracy. This makes genuinely low-latency voice agents possible — the kind where users don't perceive a pause between speaking and the AI responding.
Best-in-class accuracy at the lowest price point
Voxtral Mini Transcribe V2 achieves approximately 4% word error rate on the FLEURS benchmark across the top-10 languages and $0.003 per minute — the lowest price-to-accuracy ratio among mainstream transcription APIs. In head-to-head comparisons, it outperforms GPT-4o mini Transcribe, Gemini 2.5 Flash, AssemblyAI Universal, and Deepgram Nova on accuracy. It processes audio approximately 3× faster than ElevenLabs' Scribe v2 at one-fifth the cost, making it the clear choice for high-volume transcription workloads.
Speaker diarization that actually works across domains
Speaker diarization — identifying which speaker said what — has long been the weak point of commodity transcription APIs. Voxtral Mini Transcribe V2 includes first-class diarization with precise start and end timestamps. On the average diarization error rate benchmark across five English datasets (Switchboard, CallHome, AMI-IHM, AMI-SDM, SBCSAE) and the TalkBank multilingual benchmark, Voxtral sets a new state-of-the-art. The model handles overlapping speech by transcribing one speaker, which matches real-world meeting behavior better than models that skip overlapping segments entirely.
Context biasing for domain-specific vocabulary
Generic transcription models fail on proper nouns, technical terms, and product names — the words that matter most in professional transcription. Voxtral Mini Transcribe V2 allows you to provide up to 100 words or phrases that the model uses to bias its output toward correct spellings. This is particularly useful for medical transcription (drug names, anatomy), legal work (case names, statutes), and enterprise software (product names, internal jargon). Context biasing is optimized for English; multilingual support is experimental.
Open weights under Apache 2.0 for edge deployment
Voxtral Realtime ships as a 4B-parameter model with weights released under the Apache 2.0 license on Hugging Face. This makes it deployable on edge devices — useful for healthcare, legal, and defense applications where audio cannot leave the device or private network. At 4B parameters, it's small enough to run on modern embedded hardware with sufficient GPU memory. The Apache 2.0 license imposes no commercial restrictions, no usage caps, and no attribution requirements beyond standard open-source practice.
Benchmark Comparison
Word error rate (WER) from FLEURS benchmark (lower is better). Pricing as of February 2026.
| Model | WER | Price | Diarization | Open Weights |
|---|---|---|---|---|
| Voxtral Mini Transcribe V2 | ~4% | $0.003/min | Yes | No |
| Voxtral Realtime | ~4–6% | $0.006/min | No | Yes (Apache 2.0) |
| GPT-4o mini Transcribe | Higher | ~$0.006/min | No | No |
| AssemblyAI Universal | Higher | ~$0.006/min | Yes (add-on) | No |
| Deepgram Nova | Higher | ~$0.004/min | Yes (add-on) | No |
| ElevenLabs Scribe v2 | ~Voxtral V2 | $0.015/min | Yes | No |
Source: Mistral AI benchmark publication, February 2026. WER comparisons are relative rankings from Mistral's published results; check individual providers for current pricing.
Pricing
Voxtral Mini Transcribe V2
Batch$0.003 / minute
Batch transcription API. Speaker diarization, context biasing, word-level timestamps, 13 languages. Files up to 3 hours. Available now via Mistral La Plateforme and Mistral Studio audio playground.
Voxtral Realtime
Realtime$0.006 / minute
Streaming transcription API. Sub-200ms configurable latency. 13 languages. Also available as open weights (Apache 2.0) on Hugging Face for self-hosted deployment.
Self-hosted (Apache 2.0)
Open WeightsCompute cost only
Voxtral Realtime open weights on Hugging Face. 4B parameters — runs on edge hardware. Apache 2.0 license, no commercial restrictions. Best for HIPAA/GDPR-sensitive deployments.
Use Cases
- →Voice agents: Connect Voxtral Realtime to an LLM + TTS pipeline for sub-200ms voice interfaces. At 480ms delay, accuracy is within 1–2% of the batch model.
- →Meeting intelligence: Voxtral Mini Transcribe V2 with diarization handles multilingual meetings with speaker attribution at $0.003/min — cheap enough to transcribe your full meeting library.
- →Contact center automation: Real-time transcription with speaker diarization enables live sentiment analysis, agent suggestions, and CRM auto-fill during calls.
- →HIPAA-compliant transcription: Voxtral Realtime open weights run fully on-prem with no audio leaving your infrastructure — the cleanest compliance posture for healthcare audio.
- →Media and broadcast: Multilingual live subtitling at configurable latency with context biasing for proper nouns and technical terminology.
Frequently Asked Questions
What is Voxtral Transcribe 2?
Voxtral Transcribe 2 is Mistral AI's next-generation speech-to-text model family, released February 4, 2026. It includes two models: Voxtral Mini Transcribe V2, optimized for batch transcription with the lowest word error rate and $0.003/min pricing, and Voxtral Realtime, purpose-built for live streaming applications with latency configurable down to sub-200ms. Both support 13 languages and speaker diarization (batch model). Voxtral Realtime is available as open weights under Apache 2.0.
How does Voxtral Transcribe 2 compare to Whisper and AssemblyAI?
Voxtral Mini Transcribe V2 outperforms AssemblyAI Universal, Deepgram Nova, and GPT-4o mini Transcribe on word error rate at a lower price point ($0.003/min). Against OpenAI Whisper (API), Voxtral offers first-class speaker diarization and context biasing out of the box, which Whisper lacks. On multilingual benchmarks (FLEURS), Voxtral's non-English performance significantly outpaces competitors according to Mistral's published results.
What is Voxtral Realtime and how does it achieve sub-200ms latency?
Voxtral Realtime uses a novel streaming architecture designed from scratch for live transcription — it does not work by chunking audio and processing offline. This means it can transcribe audio as it arrives, with latency configurable down to sub-200ms. At 480ms delay, it stays within 1–2% word error rate of the batch model, making it suitable for voice agents and real-time applications. It's a 4B-parameter model available under Apache 2.0.
What is context biasing in Voxtral?
Context biasing lets you provide up to 100 words or phrases that guide Voxtral toward correct spellings of domain-specific vocabulary — names, medical terms, product names, legal terms. This significantly reduces errors on the words that matter most in professional transcription. Context biasing is currently optimized for English; support for other languages is experimental.
Is Voxtral Transcribe 2 HIPAA and GDPR compliant?
Yes. Both Voxtral models support GDPR and HIPAA-compliant deployments through secure on-premise or private cloud setups. Voxtral Realtime's Apache 2.0 open weights make full on-prem deployment possible without sending audio to a third-party API — which is the highest-confidence compliance posture for sensitive audio.
Where can I try Voxtral Transcribe 2 without writing code?
Mistral launched an audio playground in Mistral Studio (studio.mistral.ai) that runs Voxtral Transcribe 2. You can upload up to 10 audio files (mp3, wav, m4a, flac, ogg — up to 1GB each), toggle diarization, choose timestamp granularity, and add context bias terms. No API key required for the Studio trial.
Verdict
Voxtral Transcribe 2 sets a new price-performance baseline for speech-to-text APIs. Voxtral Mini Transcribe V2 beats the established players — AssemblyAI, Deepgram, GPT-4o mini, Gemini 2.5 Flash — on accuracy at $0.003/min. Voxtral Realtime's sub-200ms streaming architecture opens use cases that were previously impractical with batch-adapted models.
For teams already using Mistral's LLM API, adding STT under the same key is zero additional vendor overhead. For teams switching from AssemblyAI or Deepgram, the accuracy improvement at lower cost is a straightforward win. The Apache 2.0 open weights on Voxtral Realtime make it the default choice for any deployment where audio can't leave your infrastructure.