Speaker ID

Pick a recording on the left.

Labelled speech per person and audio band. Aim for ≥ 60 s per band across several recordings before trusting a voice.

model max wrong IDs

Pick a labelled recording. Soniox transcribes it once (exactly as call_hub's batch path asks: stt-async-v5, speaker diarization), then every word is checked at its own timestamp against your labels — silence can't count. Each system's speakers are mapped to people so that the most words agree.

Speaker ID

Speakers — assign each diarized speaker to a person; that labels all their segments

Labels — drag on the waveform to label a span by hand

Transcript — each word: your label · what Soniox says · what pyannote says. Click a word to play from there.

Who is this?

Assist token