Mistral AI Releases Voxtral 3 Real-Time Voice Agent Stack
The French AI company has released Voxtral 3, coupling low-latency transcription and expressive speech models with a LiveKit-powered voice agent framework.
In short
- Mistral launched Voxtral 3 on October 20, 2026, combining speech-to-text, text-to-speech, and an agent SDK built on LiveKit.
- Voxtral Small Transcribe 3 posted a 4.1% word error rate on FLEURS 45 in Mistral's published benchmark results, while Voxtral Mini Transcribe Realtime 3 reaches sub-500ms latency at $0.008 per minute.
- Voxtral Speech 3 generates expressive audio across 35+ languages and 50+ dialects, priced at $13 per million characters.
- The full stack supports self-hosting and EU-hosted APIs, providing data sovereignty and retention controls for regulated industries.
On October 20, 2026, Mistral AI launched Voxtral 3, a full-stack voice suite designed to build conversational AI agents that listen, reason, and speak in real time. The release combines three core components: multilingual speech recognition models, an expressive text-to-speech engine, and an agent software development kit built on the open-source LiveKit framework. According to Mistral's announcement, the software unites audio ingestion, reasoning, tool execution, and voice output into a unified streaming pipeline.
The launch expands the Paris-based laboratory beyond text and multimodal reasoning into end-to-end spoken interaction. Arthur Mensch stated that voice will become the primary interface for AI systems, noting that the company co-locates its voice models physically close to its reasoning models. The models are available immediately through Mistral's European Union, United States, and global API endpoints, with self-deployment options available for private enterprise infrastructure.
Multilingual Speech Recognition Models and Architecture
The listening tier of the Voxtral 3 family includes three distinct speech-to-text models tailored for batch transcription and live streaming, as detailed by Aipolix and Guillaume Lample. Transcription capabilities span more than 50 languages across the family, including all 24 official European Union languages alongside low-resource Asian and Indic languages.
At the top of the recognition lineup sits Voxtral Small Transcribe 3. In benchmarks published by Mistral on the 45-language FLEURS benchmark, Voxtral Small Transcribe 3 recorded a 4.1% word error rate when the target language was specified and 4.0% when the language was inferred automatically. On the same FLEURS test, Mistral reported that OpenAI's Whisper large-v3 achieved error rates of 16.0% and 14.9%, respectively, while ElevenLabs Scribe v2 registered 4.9% across both settings. Independent evaluations from Artificial Analysis ranked Voxtral Small Transcribe 3 third on its Speech to Text leaderboard with an Artificial Analysis WER Index of 2.1%, placing it ahead of Scribe v2 at 2.2%.
For low-cost batch processing, Mistral introduced Voxtral Mini Transcribe 3, a 4-billion-parameter model that achieved fifth place on the Artificial Analysis leaderboard with a 2.2% index. For interactive environments, Voxtral Mini Transcribe Realtime 3 provides streaming transcription in more than 35 languages at sub-500ms latency. According to benchmark figures from Artificial Analysis reported via Digg, the real-time model posted a 3.3% final-transcript error rate at a 960-millisecond setting and 3.6% at 480 milliseconds.
Both batch and streaming models incorporate speaker diarization to differentiate who is speaking; in Mistral's reported evaluations on public diarization benchmarks, Voxtral Small Transcribe 3 achieved an 18.4% concatenated minimum permutation word error rate, while Voxtral Mini scored 18.6%. The models also include context biasing, which allows developers to provide lists of specialist vocabulary, product names, or acronyms to improve recognition accuracy without fine-tuning weights. In Mistral's reported evaluations on the Earnings 22 dataset, keyphrase recall reached 94.6% with contextual prompting.
Expressive Speech Synthesis and Voice Adaptation
On the generation side, Mistral released Voxtral Speech 3, also designated as Voxtral Mini TTS 3. The engine supports speech synthesis across more than 35 languages and 50 dialects, offering voice presets in 18 primary languages including Catalan, Modern Standard and Emirati Arabic, Mandarin Chinese, and Brazilian Portuguese. Technical documentation shows the engine adapts rhythm, pacing, and emotional delivery based on conversational context, while Digg reported emotion and pace can be specified via explicit inline tags such as [emotion=sad][pace=slow].
In evaluations conducted by Artificial Analysis, Voxtral 3 TTS scored an Elo rating of 1,218 across 1,599 appearances in the Provider Voice Arena, ranking twelfth out of 98 models tested. In pronunciation robustness tests, the model scored 84.4%, ranking thirteenth of 32 tested systems, while reaching 93.9% accuracy on standalone terms.
For enterprise branding, Mistral added a feature called Pro Voice adaptation. According to the company's technical specifications, the tool can reproduce an individual speaker's vocal characteristics, emotional expressiveness, and speaking style across multiple supported languages using 30 minutes of recorded sample audio.
Real-Time Agent SDK and Tool Calling
The software development kit provides the orchestration layer connecting the voice models to backend computing infrastructure. Built upon LiveKit, the SDK establishes full-duplex streaming connections over WebRTC for web and mobile applications, as well as SIP protocol support for traditional telephone call centers, as detailed by Aipolix.
In operation, Voxtral Realtime Transcribe processes inbound caller audio while acoustic and semantic end-of-turn detection determines when the speaker has finished. That text feeds into an underlying reasoning model, such as Mistral Small, which can query external databases, call internal APIs, or execute web searches. The generated response streams directly into Voxtral Speech 3. Mistral states that complete conversation turns execute in under one second. The SDK also handles barge-in interruptions, halting model speech output immediately when the user speaks.
Enterprise Pricing and Sovereign Deployment
Mistral has structured Voxtral 3 pricing across separate per-minute and per-character rates. Voxtral Mini Transcribe 3 costs $0.0035 per minute for batch audio, while Voxtral Small Transcribe 3 costs $0.004 per minute. The streaming model, Voxtral Mini Transcribe Realtime 3, is priced at $0.008 per minute. For voice synthesis, Voxtral Speech 3 costs $13 per 1 million characters. Published rate comparisons show this pricing sits below Gemini 3.8 Flash TTS at $16.50 per million characters and ElevenLabs Eleven v4 Turbo at $40 per million characters.
A primary commercial emphasis for the release is data sovereignty. Mistral allows organizations to process interactions entirely within its European Union cloud facilities or deploy the models directly onto on-premises infrastructure. Under both options, customers retain administrative authority over whether audio recordings or transcripts are retained, addressing compliance restrictions for public sector bodies, financial institutions, and healthcare providers.
Questions people ask
What models are included in Mistral's Voxtral 3 release?
Voxtral 3 includes three transcription models (Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, and Voxtral Mini Transcribe Realtime 3), a speech synthesis model (Voxtral Speech 3 / Voxtral Mini TTS 3), and a LiveKit-based voice agent SDK.
How does Voxtral Small Transcribe 3 perform on speech benchmarks?
Mistral reported a 4.1% word error rate on the FLEURS 45 multilingual benchmark with language specified and 4.0% when inferred. Artificial Analysis ranked the model third on its Speech to Text leaderboard with an index score of 2.1%.
How much does it cost to run Voxtral 3 models via API?
Batch transcription costs $0.0035 per minute for Mini and $0.004 per minute for Small. Real-time streaming transcription costs $0.008 per minute. Text-to-speech output is priced at $13 per 1 million characters.
Where can developers deploy Voxtral 3 voice agents?
Voxtral 3 can be accessed via Mistral's EU-hosted, US, and global API endpoints, or deployed directly onto an organization's private infrastructure for complete data sovereignty.
Sources
- Voxtral 3Mistral, Oct 20, 2026, official source
- Mistral AI launches Voxtral 3, a full voice agent stack that listens, thinks, and talks in real timeThe Decoder, Oct 20, 2026
- Mistral Voxtral 3 Posts 4.1% WER on FLEURS, Beating WhisperAI Weekly, Oct 20, 2026
- Mistral AI Launches Voxtral 3 With Voice Agent SDK and 50+ Languagestechnobezz.com, Oct 20, 2026
- Mistral launches Voxtral 3, a family of transcription and text-to-speech models with a real-time voice agent SDKaipolix.com, Oct 20, 2026
- Mistral introduces Voxtral 3 voice models and a LiveKit-based agent SDK · Diggdigg.com, Oct 20, 2026
How this story was made: the newsroom picked it up from mistral.ai, the-decoder.com and Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (28 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 21, 2026 at 01:18 UTC
