Transcribe a voice memo or meeting recording privately

Why it matters where it is transcribed

The recordings people most want typed up are often the ones they least want to hand over: a voice memo with a business idea or a diary entry, a doctor's appointment, a call with a lawyer, a one-to-one with a manager, a board meeting. Most transcription services work by uploading the audio to their servers and processing it there. Even when that is done carefully, a copy of the conversation now exists somewhere you don't control.

Transcribe works the other way round. It downloads OpenAI's open-source Whisper speech model to your browser once and runs it there, on your device's own processor:

Once the model is downloaded, Transcribe works with no internet connection at all. Switching to flight mode before you start is an easy way to be certain nothing leaves the device.

Getting the recording as a file

Transcribe opens audio and video files: M4A, MP3, WAV, AAC, OGG, FLAC, MP4, MOV and WebM. The only real step is getting the recording out of the app that made it.

• iPhone voice memos. In the voice memos app, open the recording's menu, choose Share, then Save to Files. It saves as an .m4a, which Transcribe opens directly from Files.
• Android recordings. Most recorder apps have Share or Save; choose to save the file to your phone's storage, or share it to a folder you can find again.
• Video calls. A meeting recorded by a video-calling service is usually saved as an MP4 or M4A, on your computer or in your account's cloud recordings, from where you can download it. Transcribe takes the video file as it is and listens to its soundtrack.
• A dictaphone or camera. Copy the file across by cable or card reader.
• Recording now. Press Record in Transcribe and it transcribes when you stop. That recording is not kept, only the text, so if you want the audio too, record it in Voice Recorder (it saves M4A or WAV on your device) and open that file in Transcribe.

Transcribe it, step by step

1. Open the file. Drop it on the page on a computer, or tap to choose it on a phone.
2. Set the language if you know it. Auto works well on longer recordings; for a 20-second memo, choosing the language avoids a wrong guess.
3. Choose the model. Fast (Whisper base) is free and good on clear speech. Accurate (Whisper small, with Pro) is about three times slower but better with accents, noisy rooms and names.
4. Press Transcribe. Long recordings are cut at quiet moments into parts of about 30 seconds, so the text appears as it goes and Stop keeps what is already done.
5. Read it through while it plays. Tap any line's time to hear that moment and correct the words.
6. Save or send it. Export as plain text, text with times, Markdown or subtitles, copy it all, or send the text to another Yuzool app.

On a recent laptop the Fast model runs at about two to three times real time, so a 30-minute meeting takes roughly 10 to 15 minutes. Phones are slower, and Safari runs the model on a single processor core, so for long recordings a computer with Chrome or Edge is the quickest.

Meetings: speakers, names and length

A meeting is harder to transcribe than a memo. People talk over each other, sit at different distances from the microphone and use names and jargon the model has never heard.

• Speakers aren't labelled. Transcribe writes what is said with timestamps but does not tell voices apart. If it matters who said what, type the speaker's name at the start of their lines as you read through; the line menu's Split at the cursor helps where two people share a line.
• Fix names once. The model spells unfamiliar names by sound. Use Find and replace to correct "Shivon" to "Siobhan" everywhere at once. With Pro, a word list fixes the same names in every new transcript automatically.
• Mind the length. One file can be up to four hours, but very long files use a lot of memory. Up to about two hours works well on a computer; split a full-day workshop into sessions.
• Other languages. Whisper knows about 100 languages. With Pro, it can also translate speech in any of them into English text, handy for a meeting held in another language.

From transcript to useful notes

A raw transcript is long and conversational. A few minutes of tidying turns it into something people will actually read.

You can also send the transcript straight to another Yuzool app from the export sheet: to Mind Map to sort the ideas from a brainstorm into branches, or to Grammar Check to tidy spelling and punctuation before you share it.

Recording the next one well

Most transcription mistakes come from the recording, not the model. These make the biggest difference:

• A laptop at the end of the table. Put the phone or microphone in the middle, nearer the quietest speaker.
• Background noise. Air conditioning, a café, a TV in the next room and wind all cost words. Close the door.
• Talking over each other. Overlapping speech comes out as a jumble. A chair who lets one person speak at a time helps the transcript as much as the meeting.
• Tapping near the microphone. Typing, cups and pens on the table are loud to a phone lying on it. A folded cloth under it helps.
• Mumbling the important bits. When you dictate a memo, say names and numbers clearly, and spell an unusual name out loud once.