Back/Voxtral ASR (Mistral)
Open SourceApache 2.0 (open-weights)Self-hostable

Voxtral ASR (Mistral)

European open-weights ASR — French/German native, Mistral ecosystem

120ms
Latency (best case) ?
250ms
Latency (typical) ?
5%
WER (general audio) ?
$0.0030/min
Price per minute

Comparative Scores

Accuracy (WER)?7/10
Streaming latency?8/10
Multilingual6/10
Sovereignty?9/10
Price accessibility9/10
Streaming quality?7/10

Architecture

ArchitectureMistral multimodal (speech encoder + LLM decoder, open-weights)
Parameters~7B estimated
Languages30+
Self-hostable Yes
Streaming ? Yes
WER clean audio ?3%
DigiDouble
Phase 1 MVP — Stack Mistral souverain

High priority evaluation for Phase 1 MVP. Creates full-stack sovereign pipeline with Voxtral TTS + Mistral LLM. French/German native support critical for Swiss market. Very new — needs 3-month production testing before Phase 1 deployment.

Analysis

Voxtral ASR is Mistral's open-weights speech recognition model (announced March 2026). Part of the Voxtral full-stack (ASR + TTS). European open-weights with native French/German support. Creates a potential full-stack sovereign pipeline: Voxtral ASR → Mistral LLM → Voxtral TTS. Very new — production testing required.

Strengths

  • European open-weights (Apache 2.0)
  • Native French/German support
  • Full-stack Mistral ecosystem (ASR+LLM+TTS)
  • Self-hostable — full sovereignty
  • ~120ms estimated latency

Weaknesses

  • Very new (March 2026) — no production track record
  • Benchmarks pending
  • ~7B params — significant GPU requirement
  • No speaker diarization

STT Capabilities

Streaming ? Yes

Streaming-capable (announced). Part of Voxtral full-stack (ASR + TTS). 120ms estimated latency.

Diarization ? No
Custom Vocabulary No
Word Timestamps Yes
Auto Punctuation Yes
Multilingual Yes

30+ languages

Pricing

Price / minute
$0.0030
Price / hour
$0.180
Free tier
Open-weights — free self-hosted

Self-hosted: free. Mistral API: ~$0.003/min (estimated). Open-weights for sovereign deployment.

Sovereignty & Compliance

On-premise Yes

Open-weights — full on-premise. EU-origin (French company).

GDPR ? Compliant

Data residency: EU (Mistral, France). Full control if self-hosted.

On-premise Yes

Open-weights, self-hostable. EU-origin model. Mistral API also available.

Self-hosted Deployment

Open-weights, self-hostable. EU-origin model. Mistral API also available.

Strategic & Business Analysis

Voxtral ASR (Mistral) — Strategic Positioning

Beyond technical specs: where does this tool sit in the ecosystem, what are the risks and strategic implications for DigiDouble?

Voxtral ASR is Mistral AI's entry into speech recognition — EU-aligned, open-source, and naturally integrated with the Mistral LLM ecosystem. The sovereignty-first STT for teams already in the Mistral stack.

Cloud + On-premise
Lock-in risk:Low
Sovereignty fit:High
Open-source threat:Low
Pricing:Falling ↓

A. Strategic Positioning

Target customer: Developer / Enterprise — Mistral ecosystem, EU-aligned, open-source

Mistral AI's open-source STT — EU-aligned, high accuracy, low latency, with diarization, available as cloud API and self-hosted model.

B. Competitive Moat

  • Mistral AI ecosystem integration — natural fit for teams already using Mistral LLMs
  • EU-aligned company (French) — sovereignty-friendly positioning
  • Open-source + cloud API hybrid — flexibility for different deployment needs

Vulnerability: Newer entrant vs established STT players. Limited benchmark data vs Deepgram/AssemblyAI. No explicit compliance certifications.

E. Strategic Questions for DigiDouble

Sovereignty fit

French company, EU-aligned, open-source with self-hosted option. Natural sovereignty fit for DigiDouble's Swiss/EU requirements.

Build vs. Buy

Natural fit if DigiDouble already uses Mistral LLMs. Self-host for Phase 2 sovereignty. Cloud API for Phase 1 speed.

Lock-in risk

Open-source option eliminates vendor lock-in. Cloud API creates soft dependency on Mistral platform.

Roadmap alignment

Good alignment for EU-sovereign deployments. Mistral ecosystem integration reduces overall stack complexity.

Data Freshness

Updated 30 April 2026

Mistral AI announcement, Mar 26, 2026

Update note: Voxtral announced Mar 26, 2026 by Mistral AI. Apache 2.0 open-weights. Part of Voxtral full-stack (ASR+TTS). Production benchmarks not yet available. Mistral API pricing estimated.