Voxtral 3

Voxtral 3 is a generation of Mistral AI speech models released in public preview on October 20, 2026. It includes models for recorded and live speech recognition, models for text-to-speech, and a Voice Agents SDK that connects transcription, a reasoning model, speech generation, and tools. It is part of the wider Voxtral family, rather than a new name for the original 2025 models or the March 2026 TTS checkpoint.^[1][2][3][4]

The developer documentation distinguishes three transcription models and two speech-generation modes. These components have different input requirements and capabilities. In particular, a voice agent uses a separate language model for reasoning and tool calls; its capabilities do not all belong to the transcription model.^[2][13]

Models and access

Mistral's documentation lists the following identifiers. The model cards label the generation as public preview, while the audio overview describes the separate offline and realtime paths.^{[2][3][4][5][6][8][9]}

The distinction between streaming a response and running a realtime model matters. An offline transcription request can return the text of a complete recording in pieces. Offline TTS can likewise stream generated audio before the whole result is ready. Neither means that the request accepts a continuing live input stream.^[3][8]

The launch announcement offers hosted access through European, US, and global endpoints, plus private-cloud or on-premises deployment by arrangement with Mistral. It does not provide a public Voxtral 3 weight download or a reusable weight license. Availability for private deployment should not be read as a statement that this generation has publicly downloadable, openly licensed weights.^[1]

Launch pricing

The October 20 announcement gives the following US dollar prices. Transcription uses minutes of input audio, while speech generation uses characters of input text. The bundled voice-agent rates depend on the reasoning model.^{[1][4][5][6][9]}

Transcription

The language matrix lists 50 languages for both recorded-audio transcription models and 45 languages and dialects for Mini Transcribe Realtime 3. The API can detect the language or accept a two-letter ISO 639-1 language code. Coverage is model-specific; it is not a claim of equal accuracy in every supported language.^[7]

Recorded transcription supports segment timestamps, speaker labels, and context biasing for names or specialized vocabulary. It does not provide word timestamps: requesting timestamp_granularities=word returns an error. For audio question answering, summarization, or translation, the launch task guide directs developers to the separate audio-in-chat path using voxtral-small-latest, identified there as Voxtral 2.^[3][12]

Realtime transcription uses a WebSocket session. The application configures the session before sending audio, receives incremental text, and closes it after the final transcript arrives. Settings become fixed when the first audio chunk is sent; changing them requires a new session.^[11]

target_streaming_delay_ms controls how much following audio the model waits for before producing text. It accepts integers from 80 to 1,200 milliseconds, or 2,400 milliseconds, with 480 milliseconds documented as the default. A shorter delay favors responsiveness; a longer one gives the recognizer more context. This configured delay is not the same measurement as network-inclusive time to a finalized transcript or a complete agent reply.^[11][17]

Realtime speaker diarization and turn detection are both supported, but they cannot be enabled together in the same session. Speaker references require diarization. These restrictions matter for applications that need both named speakers and automatic conversational turn-taking.^[3][11][12]

Speech generation and voices

The TTS interface accepts a preset voice_id, a saved custom voice, or reference audio for a one-off voice clone. Requests can return MP3, WAV, PCM, FLAC, or Opus, and audio can stream as it is generated.^[8][10] Mistral's language matrix lists 35 languages and more than 50 dialects for voxtral-mini-tts-3-0.^[7]

Delivery controls include 34 emotion tags, slow and fast pace tags, and phoneme hints for individual words. The expressiveness parameter changes how strongly the delivery follows the voice, text, and tags. The speech-generation guide documents a range of 1.0 to 5.0 and warns that stronger delivery trades off voice similarity and stability.^[8][10]

Free-form style instructions are not available at launch, despite the announcement's broader instruction-following description. Inline emotion and pace tags and custom pronunciation are documented launch features; a general prose instruction is not interchangeable with these controls.^[8][23]

Mistral also announced Pro Voice adaptation using approximately 30 minutes of recordings. Its professional-cloning guide recommends 30 minutes to three hours of clean, single-speaker audio, distinct from the short-reference, zero-shot path. The latter accepts MP3 or WAV audio encoded directly in the request, rather than an audio URL, and can save the resulting voice for reuse within a workspace.^[1][22]

Voice agents

The Voice Agents SDK combines speech recognition, a large language model, and speech synthesis into a live conversation. It builds on LiveKit for audio transport and has Python and TypeScript bindings. The agent can receive interruptions, call application functions, search the web, and retrieve information from document libraries or workspace connectors.^[13]

A saved agent is a versioned configuration; a session is one conversation using that configuration. The configuration includes model choices, instructions, voice, tool declarations, a welcome message, and interruption behavior. Applications can pin a version or override options for a single session without modifying the saved agent.^[13]

For a browser application, the backend creates the session and returns connection credentials, keeping the Mistral API key out of client code. SIP telephony is available to enterprise customers at launch. Client-side tool handlers do not run on a phone call, so a telephone agent must use server-side tools for actions it needs during that call.^[13]

Evaluation results

Independent transcription evaluation

Artificial Analysis published the following results on October 20, 2026. Its live transcription boards still showed these error rates when checked on October 21. These are benchmark-specific word error rates, not a measurement across all supported languages.^[15][16][18]

AA-WER combines roughly eight hours of audio: AA-AgentTalk contributes 50%, English VoxPopuli recordings 25%, and Earnings22 calls 25%. Transcription variants are normalized before comparison. The streaming evaluation reports final-transcript and partial-transcript accuracy separately, and measures latency after detected speech end. A final-transcript error rate should not be presented as the quality of the first partial caption a user sees.^[17][18]

Mistral's multilingual launch comparison uses a different test: a 45-language intersection of FLEURS, with word error rate for most languages and character error rate for Japanese, Korean, Thai, and Chinese. FLEURS itself is a 102-language research dataset. Results on the selected subset are neither full-corpus results nor directly comparable to the English-focused AA-WER mixture.^[1][21]

Independent speech-generation evaluation

Artificial Analysis' October 20 announcement placed Voxtral 3 TTS twelfth in both its Provider Voice and Controlled Voice arenas. The original results were:^[15]

These are dated snapshots. The live provider comparison already showed 1,217 Elo when checked on October 21, while the pronunciation result remained 84.4%.^[20] Arena ratings are relative preference estimates fitted separately for each arena, not accuracy percentages; the two Elo values should not be compared as if they shared one scale. The controlled setup standardizes reference voices, while the provider setup uses voices associated with each model.^[19]

Pronunciation Robustness measures human judgments of highlighted English spans using one pinned US English voice. It tests names and terms, context-dependent readings, exact sequences, and shorthand. The 84.4% result does not mean that this fraction of arbitrary utterances, languages, or conversations will be error-free.^[19]

Launch limitations and documentation discrepancies

Several official pages disagreed during the October 21 review. The TTS model card still described nine languages, while the language matrix and launch announcement described 35 or more. The Mini Transcribe model card listed a 32k context, whereas the audio overview and limits guide listed a 65,536-token context shared by audio and vocabulary. These should be treated as documentation discrepancies, not combined into one inferred specification.^{[2][5][7][9][12]}

The Small Transcribe model card also advertised translation, speech Q&A, summarization, and function calling, while the audio overview said transcription only at launch and the task guide assigned audio-in-chat work to a prior-generation model. The documented launch transcription path is therefore the appropriate basis for integration claims.^[2][3][4]

Migration is not simply a model-name replacement. Existing code that requests word timestamps must switch to segment timestamps, and realtime settings must be sent before audio begins. Moving aliases can also change generation: the migration guide says voxtral-mini-tts-latest will move to Voxtral 3 within a month of general availability. An explicit versioned identifier avoids relying on that future alias change.^[14]

References

1. ^1 ^2 ^3 ^4 ^5Mistral AI. "Voxtral 3." Launch announcement, October 20, 2026. mistral.ai/...voxtral-3
2. ^1 ^2 ^3 ^4 ^5Mistral AI. "Audio." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...overview
3. ^1 ^2 ^3 ^4 ^5 ^6Mistral AI. "Speech to Text." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...overview
4. ^1 ^2 ^3 ^4Mistral AI. "Voxtral Small Transcribe 3." Model card, October 20, 2026. docs.mistral.ai/...voxtral-small-transcribe-3-0
5. ^1 ^2 ^3Mistral AI. "Voxtral Mini Transcribe 3." Model card, October 20, 2026. docs.mistral.ai/...voxtral-mini-transcribe-3-0
6. ^1 ^2Mistral AI. "Voxtral Mini Transcribe Realtime 3." Model card, October 20, 2026. docs.mistral.ai/...al-mini-transcribe-realtime-3-0
7. ^1 ^2 ^3Mistral AI. "Languages." Audio specifications, accessed October 21, 2026. docs.mistral.ai/...languages
8. ^1 ^2 ^3 ^4 ^5Mistral AI. "Text to Speech." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...overview
9. ^1 ^2 ^3Mistral AI. "Voxtral TTS 3." Model card, October 20, 2026. docs.mistral.ai/...voxtral-mini-tts-3-0
10. ^1 ^2Mistral AI. "Speech generation." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...generate-speech
11. ^1 ^2 ^3Mistral AI. "Realtime transcription." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...transcribe-in-realtime
12. ^1 ^2 ^3Mistral AI. "Limits." Audio specifications, accessed October 21, 2026. docs.mistral.ai/...limits
13. ^1 ^2 ^3 ^4Mistral AI. "Voice agents." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...overview
14. ^Mistral AI. "Migrate from Voxtral 2." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...migrate-from-voxtral-2
15. ^1 ^2Artificial Analysis. Voxtral 3 speech-model evaluation announcement, October 20, 2026. x.com/...2112587905222709432
16. ^Artificial Analysis. "Speech to Text AI Model & Provider Leaderboard." Non-streaming results, accessed October 21, 2026. artificialanalysis.ai/...non-streaming
17. ^1 ^2Artificial Analysis. "Speech to Text Benchmarking Methodology." Accessed October 21, 2026. artificialanalysis.ai/...speech-to-text
18. ^1 ^2Artificial Analysis. "Speech to Text AI Model & Provider Leaderboard." Streaming results, accessed October 21, 2026. artificialanalysis.ai/...streaming
19. ^1 ^2Artificial Analysis. "Text to Speech Benchmarking Methodology." Accessed October 21, 2026. artificialanalysis.ai/...text-to-speech
20. ^Artificial Analysis. "Best Provider Voice Text to Speech (TTS) Models." Accessed October 21, 2026. artificialanalysis.ai/...models
21. ^Conneau, Alexis, et al. "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech." arXiv:2205.12446, May 25, 2022. arxiv.org/...2205.12446
22. ^Mistral AI. "Voice cloning." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...voice-cloning
23. ^Mistral AI. "Style instructions." Developer documentation, accessed October 21, 2026. docs.mistral.ai/...style-instructions