Voxtral 3: choose a voice, then set its delivery
Mistral separates voice identity from inline emotion controls, while retaining older emotional voice variants.
The voice guide distinguishes a preset's named identifier from the ID returned when a workspace voice is created. Both go in voice_id; an unknown value produces a 404 error. With Voxtral 3 voices, the same voice takes inline emotion controls. Older Voxtral 2 presets remain accepted, with separate variants for different emotions.
For a standalone message, the speech-generation guide shows a request carrying the model, text, voice and output format. The response's audio_data must be decoded from base64 and saved as an audio file. Suggested trial: generate a short fictional order update and listen to the saved result. If delivery sounds forced, Mistral recommends lowering expressiveness; higher values trade voice similarity and stability for stronger delivery. Free-form style instructions are unavailable at launch.
The tag documentation places emotion and pace markers before the affected text. They are not spoken and apply until the next tag or the input's end. Only one tag of each type may control the same section. The documented support is English; French, German, Spanish, Italian, Portuguese and Hindi are expected but untested. An unsupported tag value can still return a normal response, so request success alone does not establish that the requested control was supported.
These are developer API operations: the audio overview requires an AI Studio key for every request. Its model table limits Small Transcribe 3 to transcription at launch; it is a separate component from speech generation.
MarkTechPost also reported on the Voxtral 3 release.
