“My order is 4128… sorry, 4182.” In customer support, a nearly perfect transcript can still point to the wrong order. Quality depends on the reply and the record the system looks up.
What Voxtral 3 brings together
On October 20, Mistral introduced Voxtral 3, speech recognition and synthesis models accompanied by a voice agent kit. An agent receives audio in chunks, can query a tool and speaks the response. Diarization helps identify who speaks when. Mistral describes several hosting options, including its European API and deployment on your own infrastructure. Check availability for the components you select.
What to check in your application
Create a set of fictional calls with spoken corrections, overlapping speech, noise and language switching. Connect a mock, read-only order service and require confirmation before looking up the reference. Log exact references, mishandled corrections and time from end of speech to first response audio by language. Check human handoff for ambiguous cases.
Measure the whole workflow. Mistral’s transcription results can inform your model choice. Your service’s latency and errors will also depend on connection quality, speech synthesis and tool responses.