Voxtral Transcribe 2 is Mistral’s latest speech‑to‑text release combining a batch model (Mini Transcribe V2) and a streaming model (Realtime). It adds sub‑200ms latency, 13‑language accuracy, diarisation, context biasing, and word‑level timestamps, with pricing from $0.003 per minute and open‑weights for Realtime under Apache 2.0.

Why Voxtral matters now

Real‑time voice is surging, and teams need transcription that is fast, multilingual and private by design. Voxtral Transcribe 2 delivers sub‑200ms streaming, competitive accuracy across 13 languages, and deploy‑anywhere flexibility — including open‑weights for edge use.

What’s new in Voxtral Transcribe 2

• Two models, one release: Mini Transcribe V2 (batch) + Realtime (streaming).

• Latency: Realtime configurable down to sub‑200ms; ~2.4s mode matches Mini V2 for subtitling‑grade accuracy.

• Languages: 13 supported (EN, ZH, HI, ES, AR, FR, PT, RU, DE, JA, KO, IT, NL).

• Open weights: Realtime under Apache 2.0 for edge/private deployments.

• Price‑performance: Mini V2 at ~$0.003/min aims for the lowest WER at the lowest price point; Realtime at ~$0.006/min.

Key capabilities for enterprises

• Speaker diarisation: Who said what and when, with labelled segments; handles most scenarios (note: overlapping speech is transcribed as a single speaker).

• Context biasing: Up to 100 terms to nudge spellings for brands, jargon and names (optimised for English).

• Word‑level timestamps: Accurate alignment for subtitles, audit trails and search.

• Noise robustness & long files: Works in tough acoustics; supports recordings up to 3 hours.

• Security & compliance: Supports GDPR/HIPAA‑compliant deployments; run on‑prem or private cloud.

How Voxtral compares

Mistral positions Mini V2 as best‑in‑class price‑performance with low WER, and Realtime as near‑offline accuracy at live latencies. The post claims outperformance vs GPT‑4o mini Transcribe (OpenAI), Gemini 2.5 Flash (Google), Assembly Universal (AssemblyAI), and Deepgram Nova (Deepgram), and ~3× faster processing than ElevenLabs Scribe v2 at roughly one‑fifth the cost — per Mistral’s benchmarks. Always confirm with your audio before switching.

Practical uses

• Meeting intelligence: Multilingual notes with diarisation for clean attribution.

• Voice agents: Realtime STT (<200ms) for natural turn‑taking with your LLM + TTS pipeline.

• Contact centres: Live guidance, CRM autofill, and sentiment while calls run.

• Broadcast & media: Low‑latency live subtitles; resilient to names and jargon using context biasing.

• Compliance: Timestamps and diarisation to support audits.

Try it now

You can test Voxtral Transcribe 2 immediately in Mistral Studio’s audio playground (upload up to 10 files, toggle diarisation, set timestamp granularity, and add bias terms), or integrate via API. Mini V2 is listed at $0.003/min; Realtime at $0.006/min; Realtime weights are on Hugging Face under Apache 2.0.

Summary

If you need fast, accurate, and controllable STT with enterprise features — and want the option to run privately — Voxtral Transcribe 2 is compelling. Start in the playground, benchmark with your own audio, then choose Mini V2 for batches or Realtime for live use.

FAQ

Is Voxtral Realtime really sub‑200ms?
Yes — the streaming architecture transcribes as audio arrives, with delay configurable down to sub‑200ms.

Which languages does it support?
Thirteen: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, Dutch.

Does it do diarisation and timestamps?
Yes — diarisation with labels and start/end times, plus word‑level timestamps for alignment.

Can I deploy it on‑prem or edge?
Yes — Realtime ships open‑weights (Apache 2.0) and both models support private/cloud setups aligned to GDPR/HIPAA.

What does it cost?
Indicative pricing: Mini V2 ~$0.003/min; Realtime ~$0.006/min (check Mistral for latest).