Voxtral 3 adds speech models with breaking migration changes
Mistral introduced Voxtral 3 on October 20, adding new transcription and speech generation models alongside a voice agent SDK. The release is available through Mistral’s API. Organizations seeking the full stack on their own infrastructure are directed to its sales team.
Existing transcription code needs changes
Mistral’s migration guide maps file transcription to voxtral-mini-transcribe-3-0, realtime transcription to voxtral-mini-transcribe-realtime-3-0 and speech generation to voxtral-mini-tts-3-0. Changing those identifiers is only part of the upgrade.
Word timestamps are no longer accepted. Timestamp requests must use segment granularity, while requesting word granularity returns HTTP 400. Displays that assume speaker segments never overlap also need updating. Custom vocabulary entries share the context window with audio, so large lists can exceed it.
Realtime sessions lock their settings when the first audio chunk arrives. Send configuration before audio starts and open a new session to change it. The realtime guide also says diarization and turn detection cannot be enabled in the same session.
Speech coverage is not yet consistent
The TTS model card labels voxtral-mini-tts-3-0 a public preview and lists nine languages. Yet the current language reference lists 35 languages and more than 50 dialects for that same identifier. Both pages were checked on October 21. They do not explain the difference, so neither count should be treated as a settled deployment guarantee. This conflict concerns speech generation, not the separately listed transcription coverage.
The launch announcement also promotes custom prompt instruction following for speech generation.
The dedicated style instructions page is more specific. It says the instruction parameter is unsupported on the current Voxtral TTS 3 checkpoint and describes intended behavior for a future release. Developers should not build a launch integration around that parameter.
Inline emotion and pace tags are documented separately. They apply from their position in the input until another tag or the end. The guide says these controls work in English and are untested in the six other languages it names. That qualification matters for multilingual voice design.
Test the whole conversation
The LiveKit based SDK connects transcription, a reasoning model and speech generation. Phone access is limited to enterprise customers at launch. Its guide says phone agents must use server side tools because application side functions cannot run in those calls. Browser integrations instead create sessions on a backend so API keys stay off the page.
The latency guide separates model processing time from the delay a listener experiences. Measure from the end of a user’s turn to the useful reply, including tool calls and playback. A short filler sound can cover a pause without making the answer arrive sooner.
Before switching traffic, replay representative recordings and check text, speaker labels, interruptions and the actual voice output. ByteForward has not run Voxtral 3 or measured its performance. Our Grok transcription migration report illustrates why an API transition also needs checks of the output an application consumes.
The archival microphone photograph is by Zzubnik via Wikimedia Commons and dates to May 19, 2008. Its creator released it into the public domain. This illustrates audio equipment and does not depict Mistral hardware.
