Review the record

Choose the engine. Correct the result.

Transcription and speaker detection are configurable starting points, not a locked final record. Compare the processing paths, then edit text and speaker decisions in place.

Three selectable engines Editable transcript Reversible speaker changes

Compare transcription engines

Engine Where it runs Live text Speaker handling
WhisperKit On Mac No FluidAudio, then configured-model fallback
Apple Speech On Mac No FluidAudio, then configured-model fallback
AWS Transcribe AWS cloud Yes, during live capture Native AWS speaker labels

Use WhisperKit

  • Transcription runs locally through Core ML on Apple Silicon.
  • Choose Tiny, Base, Small, Medium, or Large in Settings > Transcription.
  • Tiny is fastest with the lowest expected accuracy; Large is slowest with the highest expected accuracy.
  • Base or Small is the app's recommendation for most users.
  • The selected model downloads when needed and is reused later.

WhisperKit does not produce live text while recording. Meeting Maid transcribes the finalized file after capture stops, then runs speaker analysis.

Use Apple Speech

  • Recognition uses the built-in macOS Speech framework.
  • Meeting Maid requires on-device recognition.
  • No app-managed transcription model download is required.
  • Like WhisperKit, transcription starts after the recording is finalized and speaker turns come from FluidAudio.

Use AWS Transcribe

Select AWS Transcribe in Settings > Transcription. The profile and region come from the shared connection in Settings > AWS. Current streaming, file-streaming, and S3 batch requests use US English (en-US).

Live recordings

Microphone and system audio are mixed into one AWS stream. Partial and final text can appear in the menu bar and active meeting while recording, with native speaker labels.

Finalized files without an S3 bucket

Imports, recovery files, and other file-based work use paced AWS streaming when no bucket is configured. Audio streams at up to roughly eight times real time, subject to network and service conditions.

Finalized files with an S3 bucket

When a bucket is configured, every file-based AWS pipeline uses asynchronous batch transcription. This includes imports and recovered recordings. Live capture still uses streaming; S3 batch is a file mode, not a fourth transcription engine.

  • AWS batch transcription supports files up to the service's four-hour batch limit.
  • Meeting Maid uploads audio and retrieves the completed transcript.
  • After successful retrieval, it attempts to delete uploaded audio and output immediately.
  • An app-created bucket receives a one-day lifecycle rule for leftovers.
  • Automatic setup is intended for a dedicated bucket because it replaces that bucket's policy and lifecycle configuration instead of merging them.

Understand local speaker analysis

WhisperKit and Apple Speech use FluidAudio to identify voice turns from the source audio. The first pipeline that needs it may download and compile multi-hundred-megabyte Core ML diarization models. FluidAudio manages that cache outside Meeting Maid's storage directory.

Fallback can send transcript text to your notes provider

If FluidAudio is unavailable or returns no usable turns, Meeting Maid can send transcript text to the configured Bedrock or Ollama Notes model for fallback labels. This fallback sends text, not source audio. A remote Ollama host or Bedrock therefore receives the transcript.

Name and correct speakers

Open the Transcript tab and choose Manage Speakers above the transcript. Generic labels remain the stable detected identities; display names can be added separately.

  • Type a display name for any detected speaker.
  • Quick-pick an unused attendee from the matched invitation.
  • Quick-pick a person from Settings > Meeting Context.
  • Click a speaker label on one transcript line to reassign only that line.
  • Choose New Speaker... from a line to create another label.
  • Merge two labels when the engine split one person into several speakers.

Undo a merge or start over

  • Undo merging... reverses the newest merge. You can step back through up to eight merges until you switch recordings or quit.
  • Start over restores the originally detected labels and clears names for that meeting. It remains available after relaunch.

Start over also prevents the current invitation from immediately refilling the labels. You can still type names manually. Change the calendar match to allow automatic invitation naming to try again.

Edit transcript text

Double-click a transcript line to edit it inline. Press Return to save or Escape to cancel. Changes are saved immediately to the managed transcript file and become the text used by later note generations and exports.

Review with synchronized playback

  • Use the audio player at the bottom of the meeting detail view.
  • Press Space to play or pause.
  • Drag the seek slider or choose 0.5x, 0.75x, 1x, 1.25x, 1.5x, or 2x speed.
  • The active transcript segment highlights and scrolls into view when Auto-scroll is on.
  • Click a transcript segment to seek to its timestamp and begin playback.

To create a portable listening copy, choose Share > Export Audio. Meeting Maid proposes <safe meeting title> - Audio.m4a.