Mistral: Voxtral Mini TTS

Voxtral TTS is Mistral AI's 4B text-to-speech model for generating spoken audio from text and a voice reference. The maker's model page identifies the 2603 edition as voxtral-mini-tts-2603, while the released weight card is named Voxtral 4B TTS 2603. This entry concerns that speech-generation release, distinct from Voxtral Small audio understanding, transcription and later real-time input models.

Expressive speech and voice adaptation

The model is designed to capture a reference voice's pauses, rhythm, intonation, accent and emotional rendering. The voice acts as an instruction for how the text should sound; the maker's overview does not require separate prosody or emotion tags, or a transcript of the reference recording. The source text determines what is said, while the audio reference guides its delivery.

The overview describes adapting a voice from as little as two to three seconds of audio. The architecture description discusses voice prompts of five to twenty-five seconds. Those statements have different scopes: a short adaptation example is not a requirement to use the shortest possible reference. Choose clear speech that represents the style and accent wanted for the final result.

The released card includes twenty preset voices and adaptation to new voices. The maker's speech workflow can use a saved voice profile or a one-off reference clip. Saving a profile allows the same voice reference to be reused across separate speech requests, while a one-off clip supplies it for the current task. A profile and a generated recording are different artifacts.

Languages and cross-language delivery

The release supports English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic and Hindi, with dialect variation, cross-lingual voice transfer and code-mixing described by the maker. For the best result, the speech guide recommends matching the reference language to the text language. Cross-language transfer can preserve the reference's accent while speaking the prepared text in another supported language.

A practical narration task can use the same voice style for an explanation in different supported languages, with each text prepared for the intended listener. A voice-assistant reply can use a short, conversational script rather than reading an unedited document aloud. The text-to-speech model renders the prepared words; any translation or answer-writing stage that produces those words is a separate step.

Preparing the text

The maker's speech guide recommends verbalizable text. Write numbers and symbols in the form that should be spoken so the model does not need to choose between possible readings. A number can be a quantity, time or identifier; define the intended spoken wording before synthesis. Spell out abbreviations when pronunciation matters, using separated letters where that is the intended reading.

Avoid rich formatting, emojis and special characters that will not be rendered as spoken content. Prepare complete sentences with the punctuation and wording needed for the delivery. The guide recommends prompts below 300 words for best results. This is a quality recommendation rather than a declared hard maximum input length.

For a procedural explanation, write the action and the condition in a sentence a listener can follow. For an announcement, place the important information before optional detail. For a voice reply, retain the factual answer and remove display-only formatting or source notation that should not be spoken. These preparations make the script itself suitable for audio before the reference determines its style.

Streaming, output and duration

The maker documents basic and streaming delivery and describes native generation of up to two minutes of audio. The released weight card also lists batch inference for preparing multiple speech tasks. Its longer service generations use interleaving rather than treating one native chunk as unbounded. For longer narration, prepare coherent sections and keep their voice and text preparation consistent with the delivery workflow that handles continuation.

The released card specifies 24 kHz audio and names WAV, PCM, FLAC, MP3, AAC and Opus output formats. The speech-service guide discusses MP3 for compressed delivery, WAV for uncompressed PCM, FLAC for lossless compression and PCM or Opus for streaming. Choose a format supported by the connected workflow and suitable for its player or processing step. Format selection and streaming behavior are distinct from the voice adaptation itself.

Model architecture and practical boundaries

The announcement describes an autoregressive, flow-matching architecture with a 3.4B Transformer decoder backbone, a 390-million-parameter acoustic Transformer and a 300-million-parameter neural audio codec. The decoder predicts semantic content and the acoustic component produces speech representation, while the codec handles audio. These components explain the approximate 4B model label; they are not extra input languages or a different answer-generation model.

A practical output is the synthesized recording or audio stream associated with its prepared script and chosen voice. Keep those associations when generating several sections so the application can save, play or assemble the correct material. The model produces audio rather than a transcript, semantic-search vector or newly authored factual answer.

Preserve the 2603 speech edition, supported language and reference style. The read sources provide the native two-minute statement and the under-300-word preparation recommendation, but do not establish a separate hard maximum text-character limit or total stitched-recording duration. Use the actual text, voice and delivery requirements to define the finished audio artifact.