Speech transcription and voice era have come a good distance, however constructing end-to-end voice brokers that may pay attention, cause, work together, and deal with advanced workflows stays troublesome.
A dependable voice agent must do 4 issues very properly (and concurrently):
Listen precisely throughout languages, accents, audio system, and interruptions,
Reasondeeply about what customers imply in particular contexts,
Speak naturally with emotion and expression, and
Take motion over a number of steps by means of linked techniques.
Getting all 4 to work collectively reliably in actual time stays a problem.
To handle this, we introduce Voxtral 3: a brand new household of state-of-the-art speech-to-text and text-to-speech fashions, alongside a voice agent SDK that brings listening, pondering, device calling, and talking right into a single real-time streaming pipeline.
Voxtral 3 is constructed to create pure conversational experiences, from contact-center brokers and repair desks to digital assistants and hands-free copilots. With assist for dozens of languages, expressive speech, advanced reasoning, and deep integration with your individual instruments and knowledge, brokers constructed with Voxtral can interpret nuanced intent, cause by means of subtle requests, and execute multi-step actions throughout linked techniques.
As with different Mistral fashions, the Voxtral 3 household is designed for sovereign deployment. Available immediately by means of Mistral’s EU-hosted API or for self-deployment by yourself transport systems, it provides organizations management over the place audio is processed, how knowledge is dealt with, and whether or not conversations are retained.
Highlights
Accurate transcription throughout 50+ languages. The speech-to-text mannequin, Voxtral Small Transcribe 3, is state-of-the-art in multilingual transcription, figuring out who mentioned what (diarization), utilizing keyphrases to enhance its accuracy (context-biasing). It detects when somebody has completed talking so brokers can reply at once (end-of-turn detection).
Natural speech that adapts to the dialog. The text-to-speech mannequin, Voxtral Mini TTS 3, introduces new voices that modify their emotion, tempo, and emphasis based mostly on conversational context, with assist for 35+ languages and 50+ dialects.
Human-like intelligence at your (spoken) command. The new voice agent SDK combines listening, reasoning, device calling, and talking right into a real-time expertise. Agents can retrieve data, work together with corporate affairs techniques, and full duties throughout reside conversations over cellphone or net.
Pro Voice for distinctly correct voice adaptation. For skilled use instances the place fundamental voice cloning doesn’t suffice, Voxtral’s Pro Voice adaptation is ready to reproduce a speaker’s distinctive character, emotional vary, and talking fashion throughout languages, utilizing simply half-hour of audio.
Your voice AI, in your transport systems. Use Mistral’s EU-hosted API or deploy Voxtral 3 by yourself transport systems. Keep delicate conversations inside your surroundings and management whether or not any audio is retained.
It hears 50+ languages, with state-of-the-art accuracy.
Voxtral Transcribe 3 represents a big milestone throughout transcription, diarization (who spoke when) and context biasing capabilities. The supported checklist of languages is expanded to 50, together with 24 official EU languages and lots of low-resource Asian and Indic languages.
The Small model achieves state-of-the-art outcomes throughout diarization and multilingual transcription, together with in English, French and Spanish. It outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. The Mini variant is 4B parameters, enabling quick and correct transcriptions at a worth of $0.0035/min.
Both fashions assist context biasing, the place key phrases and free-form textual content may be given to bias the transcription and enhance recall on technical phrases. This permits them for use in manufacturing settings with domain-specific vocabulary with out the necessity for fine-tuning.
FLEURS 45 Forced vs Inferred
Word Error Rate — decrease is healthier
Macro-average WER over the intersection of 45 languages within the FLEURS benchmark supported by all seven fashions. WER is used for all languages besides Japanese, Korean, Thai, and Chinese, which use CER. Lower is healthier.
Forced = language specified; Inferred = language auto-detected. Languages: AR, BN, BG, CA, ZH, CS, DA, NL, EN, ET, FI, FR, DE, EL, GU, HI, HU, IS, ID, IT, JA, KN, KO, LV, LT, MK, MS, ML, MR, NO, FA, PL, PT, RO, RU, SK, SL, ES, SV, FIL, TE, TH, TR, UK, VI
Diarization
Concatenated Minimum Permutation WER — decrease is healthier
Macro-average concatenated minimal permutation WER (cpWER) over public diarization benchmarks: AMI-IHM, AMI-SDM, Switchboard, Switchboard CallHome, AISHELL-4. For most compatibility of API baselines, we use 30 minute variations of AMI and AISHELL. Lower is healthier.
Diarization benchmark, Concatenated Minimum Permutation Word Error Rate (cpWER).
Keyphrase Recall
With and with out contextual prompting — larger is healthier
Without Context
With Context
Voxtral Small Transcribe 3 is highlighted in model colors. Keyphrase recall on the Earnings 22 benchmarks. The context bias incorporates keyterms and a randomly sampled variety of distractors, starting from 0 to 100. Bars are superposed absolute values — the outer bar is recall with context, the inside bar is recall with out context; the seen band is the uplift from contextual prompting. Higher is healthier.
Keyphrase recall benchmark. Context = domain-specific vocabulary and speaker names equipped to the mannequin.
Artificial Analysis Benchmark
Word Error Rate — decrease is healthier
Artificial Analysis WER Index, a weighted common over VoxPopuli, Earnings-22 and AgentTalk.
Lower is healthier.
Voxtral Realtime Transcribe 3 is a realtime transcription mannequin constructed for agentic use-cases. It achieves high-quality transcriptions at sub-500ms latency in over 35 languages, together with with diarization and context biasing.
FLEURS 35
Realtime transcription Word Error Rate — decrease is healthier
Macro-average WER over 35 languages from the FLEURS benchmark. WER is used for all languages besides Japanese, Korean, and Chinese, which use CER. Lower is healthier. *Voxtral Mini Transcribe 3 is an offline baseline. The remaining fashions are realtime.
Languages: AR, BN, BG, CA, CS, NL, EN, ET, FI, FR, DE, EL, GU, HI, ID, IT, JA, KN, KO, LV, MK, MS, ML, MR, NO, FA, PL, PT, RO, RU, SK, ES, FIL, TE, TR
End-of-turn prediction is natively supported, incorporating each acoustic and semantic data to precisely predict when a consumer is completed talking. This permits for low-latency voice agent purposes, whereas minimizing false cut-off charges when the consumer is prematurely interrupted.
It solutions with the appropriate emotion, in your or your buyer’s model’s personal voice and native dialect.
Voxtral Mini TTS 3.0, accessible in batch and realtime, achieves state-of-the-art expressiveness and instruction following for 35+ languages and 50+ dialects. At $13 per 1M characters, Voxtral Speech 3 prices much less to run than comparable fashions from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.
The speech mannequin implicitly interprets the semantic that means of context higher, producing pure speech throughout a various vary of feelings and languages. New and improved options embody voice cloning from 30+ minutes of sampled audio, customized immediate instruction following (IF), contextual TTS, and a vocabulary library, all of which serve the objective of speech reliability in deployment. Variables for Explicit IF embody emotion, pacing, and emphasis.
Prompt
Let’s finish the controversy right here: it is GIF.
Let’s finish the debate right here. It’s GIF.
Let’s finish the debate right here. It’s GIF.
Non-standard phoneme trace: dʒ ɪ f
Prompt
Attends… Ce n’est pas attainable. Ce tableau vient de Reims ? Mais alors… C’est celui que ma grand-mère cherchait depuis cinquante ans !
Attends, ce n’est pas attainable? Ce tableau vient de Reims? Mais alors, c’est celui que ma grand-mère cherchait depuis 50 ans?
Attends, ce n’est pas attainable? Ce tableau viend de Reims? Mais alors? C’est celui que ma grand-mère cherchait depuis 50 ans.
Non-standard phoneme trace: ʁ e i m s
Prompt
Il nostro nuovo assistente si chiama Quivello.
Il nostro nuovo assistente si chiama Quivello.
Il nostro nuovo assistente si chiama Quivello.
Non-standard phoneme trace: tʃ i v ɛ l l o
Prompt
Hallo! Ich bin Xyra, deine persönliche Assistentin.
Hallo, ich bin Xyra, deine persönliche Assistentin.
Hallo, ich bin Xyra, deine persönliche Assistentin.
Non-standard phoneme trace: z aɪ ɹ ə
All natural-sounding speech depends on emotion, which determines its prosody, rhythm, pacing, emphasis, and transitions. Voxtral TTS natively interprets a immediate’s textual content context to find out the emotion and voice fashion, making the mannequin inherently higher at pure speech with out requiring extra directions or voice prompts.
Voice Code Bench: high quality per greenback
Voice Code Bench made by BeSimple AI is concentrated on enterprise buyer assist failure states associated to numbers and distinctive names.
Up and to the left is healthier.
Voice Code Bench (BeSimple AI): 300 English prompts, 2 voices per system (feminine and male); choose gemini-3.8-flash; costs from vendor pricing pages; generated 2026-10-20.
Pronunciation management: high quality per greenback
Share of clips the place the goal phrase is pronounced as meant, with its phonemes
Internal eval. Up and to the left is healthier. Systems with out a documented pronunciation management are listed within the knowledge, not plotted.
300 English sentences, every with one uncommon or made-up phrase (150 straightforward, 150 onerous), 2 voices per system; choose gemini-3.8-flash hears the clip with the phrase’s meant IPA; costs from vendor pricing pages; generated 2026-10-20.
Multilingual: high quality per greenback
Mean choose rating throughout 11 languages, 1 native voice per language and system
Internal eval. Up and to the left is healthier. Systems overlaying fewer than 11 languages are listed within the knowledge, not plotted.
377 prompts chosen for launch, 11 languages, unweighted imply of per-language imply choose scores, 1 to five scale; choose gemini-3.8-flash; costs from vendor pricing pages; generated 2026-10-20.
Expressive English: high quality per greenback
Mean choose rating on one_eval, 237 prompts chosen for launch
Internal eval. Up and to the left is healthier.
Expressive English (one_eval): 237 prompts chosen for launch, 2 voices per system (Voxtral TTS: Sarah and James), 1 to five scale; choose gemini-3.8-flash; costs from vendor pricing pages; generated 2026-10-20.
Illustrating Voxtral hill-climbing in expressiveness, numbers and distinctive names.
Prompt
Thank you for calling Mistral! I perceive you are having problem logging into AI Studio. Can I affirm your e mail handle? I’ve right here [email protected], is that appropriate?
Thank you for calling Mistral I perceive you are having problem logging into AI Studio. Can I affirm your e mail handle? I have right here elena.vance82 @ yahoo.com. Is that appropriate?
Thank you for calling Mistral. I perceive you are having problem logging into AI Studio. Can I affirm your e mail handle? I have right here elena.vance82 @ yahoo.com. Is that appropriate?
Prompt
Gosh that was fast, I’ve discovered a flight on your upcoming journey. How in regards to the Emirates flight at 21:00pm on 10/10 returning 10/25 at 22:00pm for £1881? Say the phrase, and I’ll e-book a window seat now.
Gosh, that was fast. I’ve discovered a flight for your upcoming journey. How about the Emirates flight at 21,000pm on 10 over 10, returning 10 over 25 at 22,000pm for £1,881? Say the phrase and I’ll e-book a window seat now.
Gosh, that was fast. I discovered a flight for your upcoming journey. How about the Emirates flight at 21pm on October tenth, returning October twenty fifth at 22pm for £1,881? Say the phrase and I’ll e-book a window seat now.
Prompt
Wow I am unable to consider you are pregnant! This is so thrilling, congratulations Corina, I’m so joyful for you!
Wow, I cannot consider you are pregnant. This is so thrilling. Congratulations, Karina. I’m so joyful for you.
Wow, I cannot consider you are pregnant. This is so thrilling. Congratulations, Karina. I’m so joyful for you.
Prompt
What do you imply my passport will not be completed this week — I’ve advised you many occasions how I must go to Paris for work. This is ridiculous. I am unable to even start to explain how infuriating this has been.
What do you imply my passport will not be completed this week? I’ve advised you a number of occasions how I want to go to Paris for work. This is ridiculous. I cannot even start to describe how infuriating this has been.
What do you imply my passport will not be completed this week? I’ve advised you a number of occasions how I want to go to Paris for work. This is ridiculous. I cannot even start to describe how infuriating this has been.
Speech sounds native in 35 languages and over 50 dialects.
Voxtral TTS helps over 35 languages and regional dialects. Our multilingual evals adhere to dialect variations in pronunciation, formal vs. casual register, spelling, verbosity, and locale conventions (time, dates, temperatures, addresses, currencies).
Voxtral 3 multilingual library
Voice Presets can be found in 18 languages: Arabic (Emirati, Modern Standard), Brazilian (Portuguese), Catalan, Chinese (Mandarin), Dutch, English (American, British), French (France), German, Greek, Hindi, Italian, Korean, Norwegian, Russian, Spanish (Spain), and Vietnamese.
The voice agent is answerable for the entire dialog.
Voxtral 3 introduces a voice agent SDK, constructed on LivePackage, that runs speech in, reasoning, and speech out as one streaming pipeline. Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has completed talking, a common mannequin—like Mistral Small—causes and calls instruments, and Voxtral Speech 3 replies within the tone of the dialog up to now. No stage waits for the one earlier than it, so every flip completes in underneath one second, and the agent stops speaking when the caller cuts in.
During a name, the agent can confirm a caller’s identification, examine a invoice, or e-book an appointment, utilizing operate calling, net search, and retrieval over your individual techniques. It solutions and locations calls over SIP, runs in browsers and apps over WebRTC, and connects to the telephony, IVR, and contact-center software program you already use.
Each stage is a separate mannequin, so you possibly can change the LLM, the voice, or the language with out rewriting the agent, and run the entire stack on Mistral’s API in Europe or by yourself transport systems.
Get began.
Voxtral 3 is accessible immediately by means of the Mistral API, hosted by means of our European, US, and international endpoints.
You can take a look at out voice brokers, audio transcription, speech era, and voice cloning within the Studio playgrounds.
Model
Price
Voxtral Mini Transcribe 3 (batch)
$0.0035/min
Voxtral Mini Transcribe Realtime 3
$0.008/min
Voxtral Small Transcribe 3
$0.0040/min
Voxtral Speech 3 (batch and realtime, identical API)
$13/1M chars
To run the complete stack on-premises or in a non-public cloud, talk to our team.
