Transcription

How transcription works

MeetingDebrief transcribes with a speech recognition model running locally on your device. No audio data leaves your computer. The model converts spoken audio into timestamped text segments. Two engines are available: Parakeet and Whisper.

Parakeet (recommended)

Parakeet TDT v3 (Multilingual) is NVIDIA's speech model and the engine MeetingDebrief suggests during setup. It covers 25 languages, including English and Dutch, with punctuation, capitalization, and word timings built in. It is about 714 MB and runs on your GPU where there is one; on Apple silicon it uses Metal. No Python environment or NVIDIA GPU is required.

Parakeet is also the engine behind dictation. Meetings and dictation share one speech pipeline, so the model stays loaded between uses and both start instantly. A dictation can interrupt a long meeting transcription, take its turn, and hand it straight back.

Control how long Parakeet stays in memory in Settings → Models under Keep the model loaded: Unload after each use, For 3 minutes (the default), or For 10 minutes. On machines with less than 16 GB of memory it gives way when summaries or chat need the room.

Whisper models

OpenAI's Whisper models are also available. Larger models produce more accurate results but require more memory and processing time:

• Tiny (English): 75 MB. Fastest, suitable for quick drafts or low-resource machines.
• Small (English): 466 MB. Higher accuracy, moderate resource usage.
• Medium (English): 1.5 GB. High accuracy, requires more memory.
• Large v3 Turbo (Multilingual): 547 MB. Covers 99+ languages.

Download models and pick the active one in Settings → Models. You can keep several installed and switch between them. Each recording remembers the model that produced it, and the new model applies to new recordings.

Live transcription

When Live transcription during recording is enabled in Settings, MeetingDebrief displays a real-time transcript while you record. This gives you a preview of what is being said during the meeting.

Live transcription may be less accurate than the final transcript generated after recording stops. The final transcription pass uses the full audio context for better results.

Automatic and manual transcription

When Auto-transcribe after recording is enabled in Settings, transcription starts automatically after a recording stops. Otherwise, open the session and click Transcribe to start manually.

Transcription runs in the background with live progress, so you can browse other meetings, take notes, or start a new recording while a long transcript is still processing. Click Cancel to stop it; a cancelled transcription never publishes a partial transcript.

Speaker labels

Transcripts label each segment by speaker. When you record both your microphone and system audio, MeetingDebrief separates Me from Other automatically, so a back-and-forth reads like a real conversation instead of one long block of text.

Speaker labels appear in the history view and are included when you copy a transcript. Naming individual speakers is on the roadmap; for now, the other side of the conversation is labelled Other.

If lines are often mislabeled, turn off Detect speakers in Settings → Audio. The change applies to transcriptions run from then on.

Vocabulary

Teach MeetingDebrief the names, acronyms, and jargon it gets wrong. Vocabulary hints your terms to the transcriber, corrects transcripts as they are saved, and explains your terms to summaries and chat. See Vocabulary.

Editing the transcript

Hover any transcript line and click the pencil to edit it in place. Press Enter to save or Esc to cancel. Edits are picked up by search and written to the Markdown mirror. When you fix a word, MeetingDebrief offers to remember the correction.

Removing phantom subtitles

Local speech models sometimes invent text during silence, such as a stray "thank you for watching", "[music]", or the same phrase repeated on a loop. MeetingDebrief recognizes and removes these known artifacts, so your transcript reflects what was actually said.

Timestamps

Each segment of the transcript includes a timestamp showing when it was spoken. Timestamps help you locate specific parts of the conversation and correlate the transcript with the audio recording.

Re-transcribing

If you switch to a different model or want to regenerate the transcript, open the session and click Re-transcribe. This replaces the existing transcript with a new one generated by the currently selected model. Summaries written from the old transcript are flagged as out of date.

Multi-language support

MeetingDebrief automatically detects the language being spoken and transcribes accordingly. The detected language appears in the session header, next to the date and duration.

When detection gets it wrong (Finnish read as Swedish, for example), use the caret on the Re-transcribe button. It opens a searchable language picker so you can force the right language, or choose Auto-detect to go back. If the selected model cannot force a language, the menu says so and offers to re-transcribe with auto-detection.

GPU acceleration

Transcription can use GPU acceleration for significantly faster processing:

• macOS: Apple Silicon devices (M1 or newer) use Metal GPU acceleration automatically.
• Windows: NVIDIA CUDA-compatible GPUs can accelerate transcription up to 10x. Download the CUDA worker from the models section in Settings to enable this.

Collapsible transcript

The transcript area can be collapsed to save screen space while you work with summaries or other parts of the session.

Copying the transcript

Click the copy button in the transcript view to copy the full transcript to your clipboard. The copied text includes timestamps and speaker labels.