How to Get a More Accurate Transcript from a Recording
What actually changes the result of automatic transcription, in order of how much it matters, and what no setting can fix.
The Scriptorium team ·
Most of the accuracy of an automatic transcript is decided before you upload: by the microphone, the room and the speakers. After that, choosing the right language helps, and so does correcting the draft. No published figure is quoted here, because it would depend on your recording.
Before you record
- Put the microphone close to the speaker. Distance and echo hurt more than almost anything.
- Cut background noise: close the window, turn off the fan, move away from the kitchen.
- One person at a time. Overlapping speech is hard for people and harder for software.
- Record in a quiet room even if the camera is the point.
When you upload
- Choose the spoken language yourself if you know it, rather than leaving it on automatic. Short or quiet clips are where detection guesses wrong.
- Use the original file rather than a re-compressed copy. See audio to text for accepted formats.
After the transcript
Search for the names and terms you know matter and correct them first. Check any number. Listen to the passages where the text looks odd; a fluent sentence can still be wrong, which is harder to spot than gibberish.
What no setting fixes
Heavy accents the recogniser has seen little of, several people talking at once, a phone call in a car, a mix of languages in one sentence, and specialist vocabulary can all produce mistakes however you upload. Treat the result as a draft and check anything you will rely on.