Tools · Meeting Notes
A meeting transcript you can listen back to and check.
Choose a recording and a small speech model on your computer writes a draft transcript. Play the recording beside it, correct what it got wrong, then save the transcript as text.
What happens
- Choose a recording. An English meeting between two people, up to 30 minutes and 250 MB, as WAV, MP3, M4A, OGG, FLAC or WebM. Your original file is not changed.
- Transcribe it. The speech model runs on your computer’s CPU. The recording stays playable while you wait, and you can stop the job.
- Listen and correct. The transcript is an editable text box beside the player, with buttons to go back or forward 5 seconds. Listen for missing or invented words, names, numbers and negatives. Speaker names are not added; you can type them in.
- Tick and save. Any edit clears the review checkbox. Once you tick that you have checked the transcript against the recording, save it as a text file.
While you work, the transcript is kept in the app’s storage on your computer, so a crash or restart can offer to bring it back. The recording itself is never copied. That stored copy is not encrypted, and it is removed when you save, discard or decline to restore it.
What it does not do
It does not write summaries or action lists, or name who is speaking. I tried suggesting actions from the transcript and it was mostly noise: about three in four suggestions were not actions, even with a small model ranking them. A hand-written action list beside the transcript distracted from checking it. So Meeting Notes stops at the transcript.
If you want a summary or actions, use Share Copy to remove the details you would rather not send, then paste the copy into a chat model of your choice.
Models used
- Whisper small.en — Systran CTranslate2 export
- English speech transcription on CPU using int8, faster-whisper 1.2.1 and CTranslate2 4.8.2.
Revision d1d751a5f8271d482d14ca55d9e2deeebbae577f. The five selected files total 486,100,128 bytes; explicit setup verifies their hashes. Normal use downloads nothing.
- model.bin — 483,545,366 bytes; SHA-256
62b2a45b05ee59acb4a5341b33ee35e041395d378d418a18acfe4c9e768ee37a. - tokenizer.json — 2,128,466 bytes; Git blob
15d7bdf9ba25718ca2504eec6a8f02bc55af0a6a. - vocabulary.txt — 422,309 bytes; Git blob
ee695b8d3e3c10d488304e04468efec4ca27554a. - config.json — 2,657 bytes; Git blob
ced3fd5bbb3e9f7664658065d2d36164712cf1d5. - Model card — 1,330 bytes; Git blob
22c0743ac43946fd981ad42117140a871e916bc7.
The upstream model and the transcription software declare MIT licences. Audio is decoded by PyAV 18.1.0, downmixed to mono and resampled to 16 kHz. Speech detection is disabled. No speaker-identification, summary or action-extraction model is active.
What has been checked
On one computer, an Apple M1 Pro, three public meetings from the AMI Meeting Corpus took 82 seconds for 17 minutes of audio, 118 seconds for 25 minutes and 132 seconds for 30 minutes. That is one run each from the command line, not through the app, and memory was not measured. My target for a 30-minute meeting was about five minutes.
Those meetings have four speakers, more than this tool is meant for. How accurate the transcripts are, and how long correcting them takes, has not been measured. Noise, accents, names and people talking over each other can all cause missing or invented words, so a transcript needs checking against the recording before you rely on it.
Availability
Not available to download yet. Meeting Notes runs in the development Kettle app on my Mac. There is no public installer or source download at present.