How to get more accurate transcripts

1. Bring the microphone close

Distance is the biggest factor. A microphone 20 cm from a mouth hears the voice; one across the room hears the room. For interviews, a clip-on or a phone placed near the speaker beats a laptop at the end of the table.

2. Use a headset on calls

It keeps the other side's voices out of your microphone, so nothing is recorded twice with an echo.

3. Pick the quietest room you can

Soft rooms (curtains, carpets, bookshelves) echo less than glass meeting rooms. Turn off fans and close windows.

4. One speaker at a time

Overlapping speech is where every system fails. A friendly "let's take turns" at the start costs nothing.

5. Say names and terms clearly once

Ask people to introduce themselves, and spell unusual names at the start. It helps you check the transcript, and it helps you set up custom words.

6. Choose the language yourself

Automatic detection listens to the opening. A short clip, a mixed-language greeting or music can mislead it.

7. Use the Accurate model for hard audio

In our tests, Whisper large-v3-turbo (Accurate) made a third to nearly half fewer mistakes than Whisper small (Standard) in every condition. With loud background noise at 5 dB SNR, Standard got 3.6% of words wrong and Accurate 2.1%.

8. Give it your names and terms as custom words

On a 118-word dictation full of unusual names, custom words took the Standard model from 6.8% of words wrong to 0.0%, and Accurate from 1.7% to 0.8%. Custom words are a Pro feature, in the import dialog and Settings › Transcription.

9. Say how many people speak

It makes speaker detection more reliable, and a transcript with the right speakers is easier to check.

10. Fix a name once, everywhere

Find and replace (Ctrl+H) corrects every instance of a misheard name in one go.

11. Slow down for the hard passages

Play a difficult stretch at 0.75x (Ctrl+[) and it often becomes clear.

12. Check what matters against the audio

All automatic transcription can mishear words and, rarely, write words that were not said, especially during silence or music. Quotes, figures, names and decisions deserve one click on the line to hear it.

• Converting the file first. Transcribe reads common formats directly; every conversion step can only lose detail.
• Turning the volume up afterwards. It makes noise louder too. A better recording beats any repair.
• Choosing the largest model for clean audio. On clear speech, Standard is already good and many times faster. Save Accurate for the recordings that need it.