# Voxtral

> Source: https://aiwiki.ai/wiki/voxtral
> Updated: 2026-07-24
> Categories: Large Language Models, Open Source AI, Speech & Audio AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Voxtral." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/voxtral
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Voxtral** is a family of speech models from [Mistral AI](https://aiwiki.ai/wiki/mistral_ai). The original open-weight speech-understanding release arrived on July 15, 2025 under the Apache 2.0 license. It ships in two sizes, Voxtral Small (24 billion parameters, for production) and Voxtral Mini (about 3 billion backbone parameters, for edge and local use), plus a transcription-tuned API endpoint called Voxtral Mini Transcribe. Unlike a pure transcription engine, Voxtral both turns speech into text and reasons over audio, handling multilingual [speech recognition](https://aiwiki.ai/wiki/speech_recognition), translation, summarization, spoken [question answering](https://aiwiki.ai/wiki/question_answering), and [function calling](https://aiwiki.ai/wiki/function_calling) directly from a voice request, with a 32,000-token context that covers audio up to roughly 40 minutes long. Mistral describes it as "open, affordable, and production-ready speech understanding for everyone," and lists hosted transcription from $0.001 per minute of audio. [1][2][3]

Voxtral marked Mistral's first serious push into open [audio models](https://aiwiki.ai/wiki/audio_models). Before it, the company was known mostly for its open and commercial text models. With Voxtral the same release pattern carried over to speech. The weights went up on [Hugging Face](https://aiwiki.ai/wiki/hugging_face), the API endpoints went live the same day, and a technical report landed on arXiv describing how the system was built and how it scored against existing models. The report opens by stating, "We present Voxtral Mini and Voxtral Small, two multimodal audio chat models." [1][3][7]

On March 23, 2026, Mistral announced **Voxtral TTS**, its first [text-to-speech model](https://aiwiki.ai/wiki/text_to_speech). Unlike the 2025 models, Voxtral TTS generates 24 kHz speech from text and a short reference voice, including zero-shot [voice cloning](https://aiwiki.ai/wiki/voice_cloning) and cross-lingual voice transfer. The hosted model is `voxtral-mini-tts-2603`, with `voxtral-mini-tts-latest` as a moving alias, while the downloadable checkpoint is `mistralai/Voxtral-4B-TTS-2603`.[11][12][14] The research paper and open checkpoint followed in the March 26 release wave. Mistral's governance page lists the model family as active.[13][18][19]

The TTS release changes the licensing picture for the family. The 2025 Voxtral Mini and Small checkpoints remain Apache 2.0, but the TTS checkpoint and its supplied reference voices use CC BY-NC 4.0, which does not permit commercial use of the downloaded weights.[12][13] Commercial hosted generation is available through Mistral's API at US$0.016 per 1,000 characters.[11][14]

## What is Voxtral?

Most open speech systems do one job well. OpenAI's [Whisper](https://aiwiki.ai/wiki/whisper), for example, is built for transcription and translation, and using it inside a larger application usually means bolting a separate [language model](https://aiwiki.ai/wiki/language_model) on top to reason about what was said.[8] Voxtral folds those two stages into one model. It listens to audio, and because its core is a Mistral language model, it can also answer questions about the audio, summarize it, or act on instructions spoken aloud. Mistral describes the result as a multimodal audio chat model rather than a transcription engine. [3]

That framing matters for the parameter sizes. A 3-billion-parameter transcription model would be large by Whisper standards, where the biggest checkpoint is around 1.5 billion parameters. Voxtral Mini is bigger because it carries a full [small language model](https://aiwiki.ai/wiki/small_language_model) inside it. Voxtral Small is larger still, built around Mistral's 24-billion-parameter text model. The trade is straightforward. You get language reasoning over speech in a single forward pass, at the cost of more weights to host. [3]

The two sizes target different deployments. Voxtral Small needs roughly 55 GB of GPU memory in bf16 or fp16, which puts it on data-center hardware. Voxtral Mini runs in under 10 GB, so it fits on a consumer card or a modest cloud instance. For teams that only want transcription and do not need the chat and reasoning layer, the Voxtral Mini Transcribe API endpoint strips the model down to its recognition behavior and is priced for high volume. [1][4][5]

## What are the Voxtral model sizes and variants?

The original Voxtral release has two open-weight sizes plus one hosted transcription endpoint. The 2026 TTS model adds a fourth offering with a different input-output direction, architecture, and license.

| Offering | Backbone | Total parameters | Form | Best for |
| --- | --- | --- | --- | --- |
| Voxtral Small | Mistral Small 3.1 (24B) | 24.3B | Open weights and API | Production transcription plus audio understanding |
| Voxtral Mini | Ministral 3B | 4.7B | Open weights and API | Local and edge deployment |
| Voxtral Mini Transcribe | Voxtral Mini, transcription-tuned | 4.7B | API only | High-volume, low-cost transcription |
| Voxtral TTS | Ministral 3 3B | About 4.1B | Open BF16 weights and API | Speech generation and voice cloning |

Voxtral Mini Transcribe is not a separate architecture. It is Voxtral Mini operated in a dedicated transcription mode and served through Mistral's API. Mistral positions it as a cheap, low-latency option that beats Whisper on accuracy while costing less than half as much per minute. The same dedicated transcription mode is available in the open weights, where it tends to maximize recognition quality and also resists instructions hidden inside the audio better than the open chat mode does. [1][6]

### Voxtral TTS

Voxtral TTS is one model marketed at roughly 4 billion parameters, not a Small and Mini pair. Its API name uses "mini," while the open checkpoint uses "4B." The model combines a 3.4B decoder, a 390M flow-matching acoustic transformer, and a 300M codec, for approximately 4.1B parameters in total.[11][13]

The hosted service accepts a one-off reference clip or a saved `voice_id` through `/v1/audio/speech`. It supports ordinary and streaming delivery.[15][16][17] The open release contains BF16 weights and 20 preset voices. Mistral recommends vLLM-Omni 0.18 or later and a GPU with at least 16 GB of memory for self-hosting.[12]

## How is Voxtral built?

Voxtral splits into three stages. An audio encoder reads the sound, an adapter compresses the encoder output, and a Mistral language model consumes the result alongside any text in the prompt. [3]

The audio encoder is the Whisper large-v3 encoder, about 640 million parameters in both Voxtral sizes. It takes a log-Mel spectrogram with 128 frequency bins and a 160-sample hop length, the same front end Whisper uses. The encoder runs over the audio and produces a sequence of acoustic features. [3]

Because that feature sequence is long, an MLP adapter sits between the encoder and the language model and downsamples it by a factor of four. The adapter is small, about 25 million parameters in the Mini and 52 million in the Small. Its job is to turn audio frames into a token stream the language model can read as if it were text. [3]

The language backbone is where the two sizes differ. Voxtral Mini is built on Ministral 3B, giving a 3.6-billion-parameter decoder and a 4.7-billion-parameter model overall.[3][5][10] Voxtral Small is built on Mistral Small 3.1, a 24-billion-parameter text model, for a 22.9-billion-parameter decoder and 24.3 billion parameters total.[3][4][9] Both retain the text abilities of their backbone, so a Voxtral model can also answer a plain text question with no audio attached.[3][4][5]

Long audio fits because of the context window. Both models carry a 32,000-token context. In its technical report Mistral states that the "32K context window enables the model to handle audio files up to 40 minutes in duration." In practice that lets Voxtral transcribe recordings up to about 30 minutes long and reason over audio up to about 40 minutes, while still leaving room for a multi-turn conversation in the same context. [1][3]

| Component | Voxtral Mini | Voxtral Small |
| --- | --- | --- |
| Audio encoder | Whisper large-v3, 640M | Whisper large-v3, 640M |
| Adapter | MLP, 25M, 4x downsample | MLP, 52M, 4x downsample |
| Language backbone | Ministral 3B, 3.6B decoder | Mistral Small 3.1, 22.9B decoder |
| Total parameters | 4.7B | 24.3B |
| Context length | 32K tokens | 32K tokens |
| GPU memory (bf16/fp16) | ~9.5 GB | ~55 GB |
| Audio limit | 30 min transcribe, 40 min understand | 30 min transcribe, 40 min understand |
| License | Apache 2.0 | Apache 2.0 |

### Voxtral TTS architecture and training

Voxtral TTS does not reuse the Whisper-encoder architecture above. Its decoder is initialized from Ministral 3 3B and consumes text tokens alongside audio tokens from a reference voice. It predicts one low-rate semantic speech token per frame autoregressively. A three-layer bidirectional flow-matching transformer predicts the corresponding acoustic tokens, and Voxtral Codec turns both streams into a waveform.[11][13]

Voxtral Codec is a 300M-parameter causal encoder-decoder trained from scratch for 24 kHz mono audio. It compresses audio to 12.5 frames per second at about 2.14 kbps. Each frame contains one semantic token from an 8,192-entry vector-quantization codebook and 36 acoustic values, each quantized to 21 levels through finite scalar quantization.[11][13] The semantic representation is distilled from a supervised Whisper recognition model. This use of Whisper for codec supervision is separate from the Whisper large-v3 encoder in the 2025 Voxtral models.[13]

Pretraining uses same-speaker audio pairs and transcripts pseudo-labeled by Voxtral Mini Transcribe. The reference and target clips need not be adjacent. Mistral freezes the text embeddings, rewrites some transcripts to vary normalized and spoken forms, and then applies [Direct Preference Optimization](https://aiwiki.ai/wiki/direct_preference_optimization) over semantic and acoustic outputs. Rejection sampling scores candidates by transcription error, speaker similarity, loudness consistency, UTMOS-v2, and language-model judging.[13] The paper does not disclose the source datasets, licenses, demographic composition, total audio hours, or per-language distribution of the main paired-speech corpus.

## What can Voxtral do?

Transcription is the base capability. Voxtral detects the spoken language automatically and writes it out, or you can pin it to a dedicated transcription mode for the cleanest output. It is natively multilingual, with strong results in English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian, among others. [1][4]

Speech translation comes from the same model. You can ask Voxtral to render spoken audio in one language as text in another, and the technical report reports high translation scores across European language pairs. [1][3]

Audio question answering and summarization use the language backbone directly. Instead of transcribing a meeting and then feeding the transcript to a separate model, you hand Voxtral the recording and ask it questions or request a structured summary. The audio and the question travel through one model together. [1][3]

Function calling from voice is the capability that leans hardest on the chat backbone. A user can speak a request, and Voxtral can emit a structured function call to trigger a backend workflow or API, with no intermediate text step that you have to parse yourself. This makes the model usable as the front end of a voice agent. [1]

Text handling survives the audio training. Because the backbone is a full Mistral model, Voxtral keeps the text-understanding behavior of Mistral Small 3.1 or Ministral 3B and can be used on text-only inputs when needed. [1][4]

### Speech generation and voice cloning

Voxtral TTS supports English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, and Hindi. It can preserve a speaker identity across languages and handle code-mixed text.[11][12][15] Mistral says a reference as short as two to three seconds can work; the paper demonstrates three-second cloning and reports that three to 25 seconds works best.[13][15]

The reference clip also guides accent, rhythm, intonation, speaking style, and emotion, so no transcript of that clip is required.[14][15] The model does not support written emotion tags or free-form style instructions. It can infer some emotion from the text, or the user can supply another clip of the same speaker performing the desired style.[13] For best pronunciation, Mistral recommends matching the reference language to the output language, spelling out abbreviations and ambiguous numbers, avoiding markdown and emoji, and keeping a request below 300 words.[16]

The model generates 24 kHz audio. Official formats include WAV, PCM, FLAC, MP3, Opus, and AAC, although the current speech-generation guide omits AAC from its shorter format table.[12][16] PCM has the lowest hosted streaming latency. The model itself natively generates about two minutes per run, while the API interleaves runs for longer outputs.[11]

## How does Voxtral perform on benchmarks?

Mistral reports that Voxtral comprehensively outperforms Whisper large-v3 on transcription, and that it beats GPT-4o mini Transcribe and Gemini 2.5 Flash across the tasks they tested. The clearest single comparison is English short-form word error rate, where lower is better. [1][3]

| Model | English short-form WER (%) |
| --- | --- |
| Voxtral Small | 1.53 |
| Whisper large-v3 | 1.84 |
| Voxtral Mini | 1.86 |
| GPT-4o mini Transcribe | 1.92 |
| Gemini 2.5 Flash | 2.97 |

On the multilingual FLEURS benchmark, Voxtral Small averages around 4 percent word error rate across the languages Mistral evaluated. The per-language figures show where it is strongest and weakest. [3]

| FLEURS language | Voxtral Small WER (%) |
| --- | --- |
| Italian | 2.62 |
| Spanish | 2.72 |
| English | 3.35 |
| German | 3.38 |
| Portuguese | 3.79 |
| French | 4.03 |
| Dutch | 4.94 |
| Hindi | 7.69 |
| Arabic | 13.44 |

For speech translation the report uses BLEU, where higher is better, and Voxtral Small posts strong scores on English-to-European-language pairs. [3]

| Translation pair | Voxtral Small BLEU |
| --- | --- |
| English to French | 57.3 |
| English to German | 47.0 |
| English to Spanish | 39.9 |
| English to Italian | 39.9 |

Audio understanding is measured with spoken-question variants of standard text benchmarks. Voxtral Small reaches 88.4 percent on OpenBook QA, 74.3 percent on MMLU, 71.7 percent on a Llama-style QA set, and 89.7 percent on GSM8K, which shows the backbone's reasoning carrying through to audio inputs. [3]

The report also gives a multilingual transcription summary. Averaged across FLEURS, Mozilla Common Voice, and Multilingual LibriSpeech, Voxtral Small lands a mean word error rate near 6.6 percent and Voxtral Mini near 7.1 percent, both ahead of the open and closed baselines Mistral compared against. [3][4][5]

### Voxtral TTS evaluations

Mistral evaluated Voxtral TTS against [ElevenLabs](https://aiwiki.ai/wiki/elevenlabs) v3 and Flash v2.5 using SEED-TTS and nine-language MiniMax-TTS test material. Automatic metrics included word error rate measured by Voxtral Mini Transcribe v2, UTMOS-v2, and ECAPA-TDNN speaker-embedding similarity.[13] Voxtral had the highest published speaker-similarity scores, but ElevenLabs Flash performed better on most automatic transcription and UTMOS comparisons. The paper also reports that ElevenLabs v3 was stronger in parts of the human emotion-steering evaluation.[13]

For flagship voices, Mistral used 77 prompts, including 66 with an expected emotion, and three native annotators of the target dialect per comparison. Audio was resampled to 24 kHz WAV. Voxtral's win rate excluding ties was 58.3 percent against ElevenLabs Flash v2.5 for implicit emotion, but 37.1 percent against [Gemini 2.5 Flash](https://aiwiki.ai/wiki/gemini_2_5_flash) TTS. With explicit emotion steering, where Voxtral used an expressive reference instead of a text instruction, its win rate was 51.0 percent against ElevenLabs v3 and 35.4 percent against Gemini.[13]

The separate zero-shot cloning evaluation used two speakers per language, 60 text prompts, and three native annotators per pair. Annotators judged likeness, naturalness, and expressivity. Voxtral's micro-average win rate against ElevenLabs Flash v2.5 was 68.4 percent, ranging from 49.4 percent in Dutch to 87.8 percent in Spanish.[11][13] The 68.4 percent figure is a Mistral-run pairwise preference result for that setup, not an independent or universal quality score. Independent launch reporting confirmed the release and capabilities but did not publish a separate reproducible audio evaluation.[19]

Mistral's latency benchmark used a single NVIDIA H200, vLLM 0.18, a 500-character input, and a ten-second reference. At concurrency one, it reported 70 ms of model latency, a standard real-time factor of 0.103, and 119 characters per second per GPU. At concurrency 32, latency was 552 ms and throughput was 1,431 characters per second per GPU.[12][13] Current API documentation separates that model time from network and encoding overhead, giving approximately 0.8 seconds to first audio for PCM and three seconds for MP3.[15]

## Is Voxtral open source, and how much does it cost?

Both open models ship under Apache 2.0, which allows commercial use, modification, and redistribution. The weights for Voxtral Small and Voxtral Mini are on Hugging Face, and the same models are reachable through Mistral's API. [1][4][5]

On the hosted side, Mistral lists API pricing starting at $0.001 per minute of audio. The company frames Voxtral Mini Transcribe as costing less than half what comparable commercial transcription services charge while still beating Whisper on accuracy, and it says Voxtral Small matches ElevenLabs Scribe on quality, again for less than half the price. Voxtral is also wired into [Le Chat](https://aiwiki.ai/wiki/le_chat), Mistral's assistant, through a voice mode on web and mobile. [1]

### Voxtral TTS access and licensing

The hosted TTS model is available in Mistral Studio, Le Chat, and the Mistral API at `/v1/audio/speech`. API pricing is US$0.016 per 1,000 characters, or US$16 per million.[11][14] Requests can use one of Mistral's voices, a saved custom voice profile, or a one-off reference recording.[15][17]

The downloadable `mistralai/Voxtral-4B-TTS-2603` checkpoint contains about 8 GB of BF16 weights and is served locally with [vLLM](https://aiwiki.ai/wiki/vllm)-Omni.[12] Mistral documents at least 16 GB of GPU memory. Unlike the Apache-licensed 2025 checkpoints, however, Voxtral TTS inherits CC BY-NC 4.0 from its distributed reference voices. It is therefore an open-weight, noncommercial release rather than open source under the Open Source Definition.[11][12][13] Mistral identifies EARS, CML-TTS, IndicVoices-R, and Arabic Natural Audio as sources of the supplied reference voices, but does not provide the same corpus-level provenance for the model's main pretraining data.[12]

## What is Voxtral used for?

The split between the two sizes maps onto real deployments. Voxtral Mini and the Mini Transcribe endpoint suit high-volume transcription pipelines, call-center logging, captioning, and on-device or edge use where a small footprint matters. Voxtral Small fits applications that need to reason about audio, such as meeting summarization, voice search, spoken question answering over recordings, and voice agents that trigger actions through function calls. Because the models keep their text abilities, a single Voxtral deployment can also serve text requests, which simplifies a stack that would otherwise run separate speech and language models. [1][3]

Voxtral TTS supplies the output half of a voice pipeline. It can turn an agent's text response into streaming speech for customer support, accessibility, narration, in-vehicle guidance, and translation. For speech-to-speech use, Mistral presents it as a cascade with Voxtral Transcribe or another recognition model rather than as a single end-to-end conversational model.[11] Cross-lingual voice transfer can preserve the selected speaker's accent and identity in the output language, subject to consent and disclosure requirements.[11][17]

## What are Voxtral's limitations?

Voxtral inherits the limits of its parts. Accuracy drops on lower-resource languages, as the Arabic and Hindi FLEURS figures show next to the European numbers. The audio length cap of roughly 30 to 40 minutes means very long recordings have to be chunked before processing, and the 32K context is shared between audio and any conversation, so heavy use of one leaves less room for the other. [3]

There is also a security wrinkle. Because the open chat models follow instructions, they can follow instructions spoken inside the audio itself, which is a prompt-injection risk for applications that feed in untrusted recordings. Mistral's dedicated transcription mode and the Mini Transcribe endpoint are more resistant to this, since they are tuned to transcribe rather than to obey, so safety-sensitive transcription jobs are better served by those paths than by the open chat model. [3][6]

### Voxtral TTS limitations and safeguards

Voxtral TTS covers nine languages, not arbitrary multilingual speech, and Mistral recommends that the reference and generated language match for the best result. Cross-lingual transfer can carry the reference accent into the output.[15][16] The model has no text-based emotion or prosody controls comparable to systems that accept style instructions or emotion tags. It relies on the reference voice and implicit cues in the text.[13] Native output is limited to about two minutes per generation, and the API's longer outputs depend on interleaving.[11]

The paper's evaluation also shows uneven behavior. Dutch voice cloning fell just below parity with ElevenLabs Flash v2.5, Gemini 2.5 Flash TTS led the flagship-voice emotion comparisons, and ElevenLabs Flash led most automatic WER and UTMOS results.[13] Mistral's tests are useful controlled comparisons but were not independently administered.

For hosted use, Mistral applies text-input content moderation and may reject a request with HTTP 403.[16] Its voice policy prohibits impersonation, cloning without explicit consent, fraud, deception, misinformation, privacy-invasive content, and other unlawful or harmful use. It also requires disclosure of generated speech where the law requires it.[17] The published documentation does not describe automated speaker-identity or consent verification, an audio watermark, content credentials, or a detector for Voxtral-generated audio.[12][13][16][17] Hosted moderation is not embedded in the downloadable checkpoint, so a self-hosted operator is responsible for access controls, consent, provenance, and disclosure.

Three-second voice cloning and cross-lingual transfer make consent failures especially consequential. The CC BY-NC license limits commercial reuse of the released weights but does not itself prevent impersonation or deceptive noncommercial use.[12][17]

## References

1. Mistral AI. "Voxtral." Mistral AI News, July 15, 2025. https://mistral.ai/news/voxtral
2. Mistral AI. "Voxtral models." Mistral AI Documentation. https://docs.mistral.ai/getting-started/models/models_overview/
3. Liu, Alexander H., et al. "Voxtral." arXiv preprint arXiv:2507.13264, July 2025. https://arxiv.org/abs/2507.13264
4. Mistral AI. "Voxtral-Small-24B-2507." Hugging Face. https://huggingface.co/mistralai/Voxtral-Small-24B-2507
5. Mistral AI. "Voxtral-Mini-3B-2507." Hugging Face. https://huggingface.co/mistralai/Voxtral-Mini-3B-2507
6. Willison, Simon. "Voxtral." Simon Willison's Weblog, July 16, 2025. https://simonwillison.net/2025/Jul/16/voxtral/
7. Mistral AI. "Voxtral technical report (HTML)." arXiv, July 2025. https://arxiv.org/html/2507.13264v1
8. OpenAI. "Introducing Whisper." OpenAI, September 21, 2022. https://openai.com/index/whisper/
9. Mistral AI. "Mistral Small 3." Mistral AI News, January 30, 2025. https://mistral.ai/news/mistral-small-3
10. Mistral AI. "Un Ministral, des Ministraux." Mistral AI News, October 16, 2024. https://mistral.ai/news/ministraux
11. Mistral AI. "Speaking of Voxtral." Mistral AI News, March 23, 2026. https://mistral.ai/news/voxtral-tts/
12. Mistral AI. "Voxtral-4B-TTS-2603." Hugging Face model card, March 2026. https://huggingface.co/mistralai/Voxtral-4B-TTS-2603
13. Liu, Alexander H., et al. "Voxtral TTS." arXiv:2603.25551v2, April 6, 2026. https://arxiv.org/html/2603.25551
14. Mistral AI. "Voxtral TTS." Model card, March 23, 2026. https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03
15. Mistral AI. "Text to Speech." Documentation. https://docs.mistral.ai/studio-api/audio/text_to_speech
16. Mistral AI. "Speech Generation." Documentation. https://docs.mistral.ai/studio-api/audio/text_to_speech/speech
17. Mistral AI. "Voices." Documentation. https://docs.mistral.ai/studio-api/audio/text_to_speech/voices
18. Mistral AI. "Voxtral Mini TTS." AI Governance model lifecycle. https://legal.mistral.ai/ai-governance/models/voxtral-mini-tts
19. Ivan Mehta. "Mistral releases a new open source model for speech generation." TechCrunch, March 26, 2026. https://techcrunch.com/2026/03/26/mistral-releases-a-new-open-source-model-for-speech-generation/