“Four binders. No, six.” Both numbers must appear in the transcript. In the materials list, only the confirmed quantity should count later on. Between these two correct results lies a small but decisive thinking task: the course of a conversation and its current working state are different things.
Mistral introduced Voxtral 3 on 20 October 2026. The documentation for the transcription models describes timestamps and speaker attribution, among other things, and lists access as a Public Preview. What is particularly interesting in practice is how the transcript, the location in time and the later evaluation work together.
What a speaker label really says
Speaker attribution can link segments spoken by the same voice. A label such as Speaker A is, to begin with, a technical assignment within the recording. It does not establish a name, a responsibility or an identity outside this context.
The timestamp, too, fulfils a limited task: it leads to a point in the original audio. A correct timestamp does not make a wrongly recognised name right. A linguistically smoothed summary, in turn, can remove the self-correction that was important for the decision.
It is therefore worth keeping three results separate: a near-verbatim transcript, the speaker and time structure, and a derived task overview. Each can be checked without silently rewriting the others.
A conversation sample of your own with a known meaning
With a second person who has agreed to it, record a short, non-sensitive scene. One person orders four binders. The other asks about an additional participant. Then the quantity is explicitly corrected to six. At the end, it remains open who will collect the binders. Use invented names and no real customer conversations.
Before any automatic processing, write a reference of your own: the wording of the change in quantity, the known changes of speaker and the open question of collection. You determine the time ranges from your actual recording. A perfect artificial audio track is not necessary for this; what matters is that you yourself reliably know the content.
First let a suitable transcription function produce only the transcript and the speaker segments. Check the change in quantity and jump back using the timestamps. Only then derive a task overview. This makes it possible to tell a recognition error apart from a wrong summary.
An agreement with a way back to the audio
For the second step, a small structure of your own is enough:
Conversation excerpt: “Four binders. No, six.”
Current working state: six binders
Evidence: time range of the explicit correction
Collection: responsibility not agreed in the conversationA suitable instruction is: “Summarise only tasks that were explicitly agreed. Take later corrections into account. For each task, give the matching point in time. Keep unclear responsibilities as an open point.” This is more precise than asking for a professional set of minutes.
Then check whether a summary has accidentally turned a non-binding suggestion into a commitment. “Maybe Lea can pick it up” is a different state of the conversation from “Lea is picking it up”. A correctly recognised word can still be given the wrong degree of commitment in the evaluation.
Subtitles need a further decision
For subtitles, you may want to shorten repetitions and form readable lines. A materials overview, on the other hand, should condense the current agreement. Both tasks can use the same source audio but need different rules.
Keep the near-verbatim basis separate from the edited version. That way you can later trace why a line was shortened and whether the correction was preserved. Proper names, numbers and qualifications deserve a targeted second listen.
The small experiment does not test a general recognition rate or performance with arbitrary accents. It shows something immediately useful: at which stage does a sequence that was heard turn into a statement that matters for action? According to the documentation, Voxtral 3 provides building blocks for this path. A usable record emerges when their results remain linked to the original and open agreements are allowed to stay open.
Sources and status
Sources as of 25 October 2026. Voxtral 3 · Voxtral Mini Transcribe 3.
