A plain transcript of a meeting is a wall of text. What makes it readable is knowing who was talking — and, when the recording is bad, which turns you can trust. nur-stt-v1-pro returns both.

Prerequisites

Create one in Settings → API keys. Keep it server-side.
WAV, MP3, FLAC, OGG, M4A or WebM. One pass handles roughly ninety minutes; split anything longer before uploading.

Recipe

Naming the speakers

The model returns S01, S02 — it knows the voices apart, not who they belong to. Map them once and apply the mapping everywhere:
Numbering is per request. Transcribe the same meeting twice and S01 may be a different person — so store the mapping against the transcript, not against the speaker id.

Telling a bad recording from a bad transcript

Each segment carries an audibility read: whether the speaker was on mic, off mic, or faint. This is the difference between “the model got it wrong” and “nobody could have got that right”.
A single speaker in a quiet room should come back as one speaker. If a solo recording is split into two, the audio is usually the reason — a change of distance from the mic, or a long pause with room tone in it.

Watching it decode

nur-stt-v1-pro streams over SSE, so segments arrive as the file decodes instead of all at once at the end. For a ninety-minute recording that is the difference between a progress bar and a blank screen.

Going further

Audio intelligence

Diarization, audibility, and sound events in full.

Speech to text

The endpoint, and choosing between the two tiers.