How to Convert Audio to Text
Turning a recording into text: which formats work, how to handle multiple speakers, and the recording habits that decide the result.
The Scriptorium team ·
Upload the audio file to a transcription tool, wait, correct the result while listening. That is the whole process. What varies is how good the starting audio is, and that is what this article is mostly about.
The steps
- Open audio to text.
- Drop in the recording. MP3, WAV, M4A, FLAC, OGG, Opus, WMA, AIFF, AMR and CAF all work as they are.
- Set the language if you know it.
- Wait for the transcript.
- Rename the speaker labels if there is more than one voice.
- Read along with the audio and correct what is wrong.
- Export.
Free for one recording up to 15 minutes and 25 MB, with no account.
Which audio formats work
All the common ones, and there is no need to convert between them first. Compressed or uncompressed makes very little practical difference: a decent MP3 of a clear conversation transcribes about as well as the WAV it came from.
What does matter is re-compression. Exporting an already-compressed file again, especially at a low bitrate, throws away detail the recogniser was using. Upload the original recording rather than a copy you made for emailing.
Where accuracy actually comes from
Almost entirely from the recording, and hardly at all from the software. In descending order of impact:
- Whether people talk over each other. The hardest case in the field.
- How close the microphone is. A phone on the table across the room records mostly room.
- Echo. Hard surfaces smear each word into the next.
- Background noise. Traffic, cafés, fans, air conditioning.
- Specialist vocabulary. Names, jargon, case numbers, drug names.
If you can influence the recording at all, moving the microphone closer will improve your transcript more than any other single change, including buying better equipment.
Multiple speakers
When a recording has several voices, speech is grouped by speaker and each turn is labelled, so an interview reads as an exchange rather than a wall of text. You rename the labels to real names once and the change applies everywhere.
This is worked out from the sound of each voice. It is very good at noticing that the speaker changed and less reliable at counting how many distinct people there are, particularly with similar voices or a lot of crosstalk. It handles up to eight speakers.
If you control the recording and the speakers matter, record each person on a separate track. That removes the problem rather than asking software to solve it.
Common problems
The transcript is empty
There is no speech on the recording, or the file has no audio stream. Play it and check.
Wrong language detected
Detection reads the opening seconds. If the recording starts with music or silence it can commit to the wrong guess. Set the language explicitly.
One speaker labelled as two
Usually caused by a change in how the person sounds — moving closer to the microphone, or a change in tone. Merge the labels manually.
Phone call audio is rough
Expected. Phone audio is narrowband, which discards frequencies the recogniser would otherwise use. It works, just less well.