Convert audio to text

People convert files before uploading because they assume a lossless format will transcribe better. In August 2026 we tested that directly: one spoken recording, encoded as MP3, WAV, M4A, FLAC and OGG, with every version sent through the identical transcription request.

All five returned the same transcript, word for word. WAV, being uncompressed, was 2,381 KB. The MP3 of the same audio was 433 KB, roughly five and a half times smaller, and produced text that was byte-identical to the WAV. FLAC, the lossless compressed format people reach for when they want quality, landed in between at 936 KB with, again, exactly the same words.

So the format is not the variable. Converting a file before uploading it cannot improve a transcript, and it can only lose information if the conversion is lossy, which is the sort of thing our M4A test actually caught. Upload whatever you have, and if you get a choice, pick the smaller file purely because it uploads faster.

What does change the result is the recording. Microphone distance, background noise, people talking over each other and how clearly they speak account for nearly all of the variation you will see. The same test made this visible in miniature: one name in the script came back slightly wrong in every single format, because the difficulty lived in the word rather than the file.

Since the recording is what matters, the highest-value habits happen before and after transcription rather than during it. Before uploading, set the spoken language explicitly instead of relying on auto-detect, which is the single biggest avoidable source of error, and type any recurring names, company names or specialist terms into the vocabulary box, a field available on every plan including free.

Afterwards, read the names first. Speech recognition is most reliable on ordinary conversational words and least reliable on proper nouns, so corrections cluster there rather than being spread evenly through the text. If a name recurs, fix it once and search for the other spellings.

For anything longer than a few minutes, the synced editor is the part that saves real time: clicking any line replays that exact moment, so verifying a quote or a figure is a two-second job rather than a hunt through the audio. That is the difference between a transcript you trust and one you have to listen through again.