Meeting

Meeting — who said what, fully on-device

Speaker-attributed transcription: Streaming Sortformer v2 (up to 4 speakers) or Nemotron-3-Diarization (up to 8) segments the clip into speaker turns ("who spoke when"), then the on-device ASR you pick (Whisper / Qwen3-ASR / Parakeet, or Apple's) transcribes each turn. Two models behind one call:

import CoreAIKit

let meeting = try await MeetingTranscriber(asr: "whisper-large-v3-turbo")
let transcript = try await meeting.transcribe(samples: try AudioFile.pcm16kMono(url))
print(transcript.text)
// Speaker 1 [0.3–4.1s]: With her white paint and her scarlet smokestack…
// Speaker 2 [4.6–6.3s]: …

Run it

cd Examples/Meeting
swift run meeting-cli --audio meeting.wav                          # Whisper turns
swift run meeting-cli --audio meeting.wav --asr parakeet-tdt-0.6b-v3
swift run meeting-cli --audio meeting.wav --diarizer nemotron-3-diarization --asr system --locale en-US
swift run meeting-cli --list-models                                # what's in the catalog

First run downloads the diarizer (226 MB Sortformer, 200 MB Nemotron-3-Diarization) and the ASR; --asr system is Apple's transcriber, which the app does not download. Afterwards everything is local — airplane mode works. Audio is anything AVFoundation reads (wav/m4a/mp3/…), resampled to 16 kHz mono.

diarize-gate checks the diarizer alone against a reference's per-frame probabilities: agreement at 0.5, the turns, and the wall time. With the model zoo's golden files for Nemotron-3-Diarization it reproduces the model card's agreement:

swift run -c release diarize-gate --audio clip.wav --golden clip_ll_probs.f32le

The pieces are separately usable: KitDiarizer alone answers "who spoke when" (diarize(samples:) → [SpeakerSegment]), and MeetingTranscriber(diarizer:transcriber:) composes any diarizer with any ASR you already loaded.

For a GUI take on the same pipeline, see the coreai-audio app's Transcribe tab ("Diarize — who said what") in the zoo repo.