A small macOS app that turns a podcast, interview or meeting recording into a transcript where every line is assigned to a speaker. Drop an audio or video file on the window and you get a speaker timeline, a clickable transcript and exports for subtitles and show notes. Everything runs on your Mac. No audio is uploaded.

It's built on Who, the on-device speaker diarization model from Desert Ant Labs, paired with Apple's built-in speech transcriber.

Who works out who spoke when: it returns numbered speaker turns with start and end times. It doesn't transcribe and doesn't know anyone's name. To get "who said what", the app runs two on-device models at the same time:

Each word goes to the speaker turn it overlaps most; a word in a gap goes to the nearest turn. Consecutive words from the same speaker become one paragraph.

• Drop zone or a "Choose File…" button. Takes WAV, MP3, M4A, AIFF, MP4 and MOV.

• Live progress with a running timer and two progress bars (Who labelling speakers, and transcription), each with a live "× realtime" figure.

• Speed tiles on the results screen for Who, speech to text and the whole pipeline.

• Speaker timeline: one coloured lane per speaker. Click anywhere to jump there.

• Transcript: click a line to play from that point. The line being played is highlighted and scrolls into view.

• Video panel for video files: the border takes the current speaker's colour and their name shows in the corner.

• Rename speakers: click a speaker chip or a name in the transcript, type a name and press Return. Names show up everywhere, including exports.

• Export (⌘E) with your speaker names:

  Format	Looks like	Good for
  Text with timecodes (.txt)	[00:01:23] Name: what they said	Show notes, reading, docs
  Subtitles (.srt)	Name: what they said, short cues	YouTube, Premiere, Final Cut, DaVinci
  Web subtitles (.vtt)	<v Name>what they said	HTML5 video players
  JSON (.json)	speakers, plus segments with start/end in seconds	Other tools and scripts

• Space bar plays and pauses. Copy transcript is in the ⋯ menu.

• macOS 26 on Apple Silicon
• Xcode 26 or newer (Swift 6.2+)

./build_app.sh                      # builds WhoSaidWhat.app (ad-hoc signed)
open WhoSaidWhat.app                # then drag a file on the window

# open straight into a file
open WhoSaidWhat.app --args -open /path/to/episode.mp4

# headless: write a transcript file (speed numbers go to stderr)
WhoSaidWhat.app/Contents/MacOS/WhoSaidWhat --print episode.wav > episode.txt
WhoSaidWhat.app/Contents/MacOS/WhoSaidWhat --print episode.mp4 --format srt --names "Host,Guest" > episode.srt
#   --format  txt (default) | srt | vtt | json
#   --names   comma-separated, in order of first appearance

The first run downloads Who's model (~81 MB, Core ML) and Apple's English speech model. After that everything works offline.

Package.swift                       SwiftPM package: macOS 26, depends on desert-ant-core 4.1
build_app.sh                        Builds release and bundles WhoSaidWhat.app
Sources/WhoSaidWhat/
  App.swift                         App entry point, plus the headless --print mode
  Pipeline.swift                    Audio prep, Who diarization, transcription, word/speaker alignment
  Export.swift                      Transcript rendering: TXT, SRT, VTT, JSON
  ContentView.swift                 UI: drop zone, progress, results, timeline, video panel, rename, export

• Audio prep: video files have their audio track decoded losslessly to a temporary PCM file with AVAssetReader, so there's no extra lossy step.
• Who input: the app decodes audio to 16 kHz mono itself and calls who.diarize(samples:sampleRate:), so the timing covers only the model.
• Phantom speakers: laughs or coughs can get their own label. keepingSpeakers(minimumShare: 0.02) drops speakers under 2% of the recording, and the UI hides any speaker left with no transcribed words.
• Timing: Who's number excludes the model download and audio decoding; the transcription number excludes the one-time install of Apple's speech model. Both models run concurrently, so the pipeline takes about as long as the slower one.

• Who labels up to 4 speakers per five-minute section and works best on close-miked audio (podcasts, interviews, remote calls). The transcriber is set to US English.
• Apple's transcriber can mishear names; rename speakers before exporting.
• Who can be unstable on synthetic text-to-speech voices; real voices work much better.

The app code in this repository is provided as a demo. Who and the Desert Ant SDK are under the Desert Ant Labs Source-Available License (not open source): free below 100,000 monthly active devices per platform, with unlimited inference. The SDK sends a monthly license check that contains no audio.