Voxtral 3 Explained: Models, Architecture, and Pricing

Mistral AI announced Voxtral 3 in October 2026 as a dedicated family of speech-to-text (STT) and text-to-speech (TTS) models paired with a real-time voice agent software development kit (SDK). Mistral targets contact-center agents, service desks, virtual assistants, and hands-free copilots.

The release pairs multilingual transcription models with dynamic speech generation and a modular orchestration layer. Organizations can consume the models through managed endpoints hosted in the European Union (EU), the United States (US), and global regions, or self-deploy on their own infrastructure.

What Voxtral 3 Is

Voxtral 3 is Mistral's family of speech-to-text and text-to-speech models plus a voice-agent SDK that runs listening, reasoning, tool calls and speech in one real-time pipeline, available on Mistral's EU, US and global API or on your own servers.

The product release includes distinct components tailored for specific voice processing jobs. On the listening side, Mistral provides three transcription models: Voxtral Small Transcribe 3 for multilingual speech recognition, Voxtral Mini Transcribe 3 for compact processing, and Voxtral Realtime Transcribe 3 for low-latency conversational audio. For speaking, Mistral provides Voxtral Mini TTS 3 alongside a specialized voice adaptation system named Pro Voice.

Mistral provides studio playgrounds where teams can test voice agents, transcription, speech generation, and voice cloning.

Transcription Capabilities and Benchmarks

Mistral built Voxtral Small Transcribe 3 for multilingual transcription, emphasizing domain adaptability and speaker separation.

The model includes native diarization to identify who spoke each phrase, end-of-turn detection to determine when a speaker has concluded, and context biasing. Context biasing lets operators pass specific vocabulary lists, product terms, or raw text blocks into the prompt to raise recognition accuracy on industry jargon without fine-tuning the base weights. Voxtral Small Transcribe 3 supports 50 languages, spanning all 24 official EU languages along with lower-resource Indic and Asian languages.

Mistral reports that Voxtral Small Transcribe 3 establishes state-of-the-art results on its benchmarks. Mistral evaluated the model on the FLEURS multilingual Word Error Rate (WER) dataset, diarization concatenated minimum-permutation Word Error Rate (cpWER), the Earnings-22 keyphrase recall benchmark, and the Artificial Analysis WER index. Mistral claims the model outperforms OpenAI Whisper, ElevenLabs Scribe v2, and Google Gemini 3.5 Transcribe across these tests. These are Mistral's own reported results.

For lightweight deployments, Voxtral Mini Transcribe 3 operates with 4 billion parameters. For interactive applications, Voxtral Realtime Transcribe 3 delivers latency under 500 milliseconds across more than 35 languages, preserving speaker diarization and context biasing during live audio streams.

• Context biasing: boosts domain vocabulary recall using reference text without retraining
• Speaker diarization: labels separate conversational participants automatically
• Turn detection: identifies pauses to signal when conversational turns conclude
• Language coverage: 50 languages covering European, Indic, and Asian language groups

Speech Synthesis and Voice Cloning

Voxtral Mini TTS 3 handles the output layer of the voice system, generating conversational audio that modulates emotion, pace, and emphasis according to the conversational context.

The synthesis model supports over 35 languages and more than 50 regional dialects. Mistral provides ready-to-use voice presets across 18 specific languages: American English, British English, Arabic (Modern Standard and Emirati), Brazilian Portuguese, Catalan, Mandarin Chinese, Dutch, French, German, Greek, Hindi, Italian, Korean, Norwegian, Russian, Spanish (Spain), and Vietnamese.

For custom brand voices, Mistral introduced Pro Voice. This adaptation model requires 30 minutes of reference audio to reproduce a speaker's character, emotional range, and speaking style. Once captured, the adapted voice can speak across multiple languages while preserving the speaker's vocal characteristics.

Voice Agent SDK Architecture on Live Calls

The Voxtral voice agent SDK uses a real-time streaming pipeline built on LiveKit.

During a call, no stage waits for the one before it to finish. Voxtral Mini Transcribe Realtime 3 continuously converts incoming caller audio into text while checking for end-of-turn signals. As soon as the caller pauses, an underlying large language model (LLM), such as Mistral Small, reasons and triggers necessary tool executions. Voxtral Mini TTS 3 begins generating and streaming synthesized speech back to the caller while the language model completes its response.

This zero-wait pipeline design allows full conversational turns to complete in under one second, according to Mistral. If the caller cuts in, the agent stops talking. The SDK connects to standard telephony systems and interactive voice response (IVR) platforms through Session Initiation Protocol (SIP), while running natively in web browsers and mobile applications using Web Real-Time Communication (WebRTC).

Because each stage in the SDK functions as an independent modular service, engineering teams can replace individual components without rebuilding the broader application. Developers can swap the reasoning LLM for another model, select a different voice preset, or change the language without rewriting the agent.

• Telephony integration: SIP support for existing telephony and contact centers
• Web and mobile audio: WebRTC support for browser-based and in-app voice interfaces
• External integrations: function calling, web search and retrieval over your own systems, to verify a caller's identity, check a bill or book an appointment
• Modular architecture: swapping the LLM, voice or language without rewriting the agent

Voxtral 3 Pricing and Usage Costs

Mistral publishes per-minute rates for its transcription models and its voice agent API, and a per-character rate for text-to-speech; Pro Voice pricing is not published.

Voxtral Mini Transcribe 3 (batch) costs $0.0035 per minute of processed audio, Voxtral Small Transcribe 3 costs $0.004 per minute, and Voxtral Mini Transcribe Realtime 3 costs $0.008 per minute. Voxtral Mini TTS 3 costs $13.00 per 1 million characters of text input. Mistral states that this speech synthesis rate is less expensive to operate than OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Google Gemini Flash. To confirm current figures, billing minimums, and usage tiers, verify rates directly on the Mistral pricing page.

Mistral has not stated a public price for Pro Voice. The voice agent API costs $0.04 per minute with Mistral Small 4 doing the reasoning, or $0.08 per minute with a larger model; to run the full stack on-premises or in a private cloud, Mistral says to talk to its team. Mistral has also not published information regarding free usage allowances, concurrency limits, or API rate limits.

• Voxtral Mini Transcribe 3: $0.0035 per minute of audio
• Voxtral Mini TTS 3: $13.00 per 1 million characters of generated speech
• Voxtral Small Transcribe 3: $0.004 per minute
• Voxtral Mini Transcribe Realtime 3: $0.008 per minute
• Voice agent API: $0.04 per minute with Mistral Small 4, $0.08 per minute with a larger model
• Pro Voice voice adaptation: pricing is not published by Mistral

Hosting Models and Data Retention Controls

Mistral says organizations control where audio is processed, how data is handled, and whether conversations are retained.

Organizations can consume Voxtral 3 through Mistral's managed cloud Application Programming Interface (API), routing requests through designated endpoints located in the European Union, the United States, or global server regions.

Mistral also supports self-deployment on your own infrastructure. Mistral has not stated whether the weights are openly downloadable for public use or what specific commercial license terms govern self-hosted deployments.

Voxtral 3 Compared to ElevenLabs

Comparing Voxtral 3 with ElevenLabs involves weighing Mistral's sovereign infrastructure controls and modular design against an established specialized voice platform.

ElevenLabs is an established hosted voice platform. Teams evaluating both vendors should compare current subscription plans and character allowances on the ElevenLabs pricing page against Mistral's published rates of $13 per 1 million characters for TTS and $0.0035 per minute for transcription. For deeper background on the competing platform, review the guides on ElevenLabs pricing, ElevenLabs architecture, and ElevenLabs agent workflows.

Per Mistral, Voxtral 3's differentiators are EU hosting or self-deployment, sovereign data control, and price. Mistral also lets teams swap the LLM, voice or language.

Regarding performance, Mistral states that Voxtral Small Transcribe 3 outperforms ElevenLabs Scribe v2 on multilingual word error rate benchmarks, and that Voxtral Mini TTS 3 costs less to operate than ElevenLabs Flash. These comparisons represent Mistral's published claims and have not been independently tested across live production workloads.

• Deployment: Voxtral 3 offers EU, US and global hosted endpoints or self-deployment on your own infrastructure; ElevenLabs is an established hosted platform.
• Scope: ElevenLabs offers text-to-speech, Scribe speech-to-text and ElevenLabs Agents; Mistral pairs its speech models with a voice agent SDK.

Evaluation Criteria for Production Voice Pilots

Deciding whether to pilot Voxtral 3 depends on an organization's internal engineering capacity, regulatory compliance profile, and existing telephony investments.

Who this is for: Organizations evaluating Voxtral 3 include teams that want EU, US or global hosted endpoints, organizations requiring self-deployment on their own infrastructure, and engineering teams looking for a voice agent pipeline with swappable models over SIP or WebRTC.

What would change this assessment: If Mistral publishes license terms for self-deployment, publishes Pro Voice pricing, or introduces fully managed no-code agent management dashboards, Voxtral 3 would become viable for a broader range of non-technical business teams.

At Layer3Labs, we build custom agents and enterprise integrations for business workflows. Teams considering an implementation should test live audio streams in the Mistral studio playground before committing development resources to custom telephony integration.