Transcribe MP3

MP3 is the format people convert to before uploading, usually out of habit. It is rarely necessary: 35 extensions upload here as they are, including every format a phone, a recorder or a video editor is likely to produce.

Converting costs something. A re-encode is a new lossy generation on top of whatever the source already lost, and consonants are the first thing to blur. If your original is an .m4a voice memo, a .wav from a recorder or a .mov from a phone, upload that instead — it will transcribe at least as well and usually better.

Where MP3 earns its place is as the smallest thing you can send when the upload is the bottleneck. At 128 kbps an hour is about 58 MB, against roughly 2 GB for the same hour of video.

These are medians measured on our production cluster, expressed as time per hour of audio.

Against the MP3 files uploaded here — 7,743 of them, a fifth of everything — the median runs 14 minutes 12 seconds and comes back in about 21 seconds. Nine in ten are under 65 minutes, roughly a minute and a half.

No. The model works on 16 kHz mono audio, so both are resampled to the same thing before transcription — the median MP3 uploaded here is 128 kbps and nothing about it is a handicap.

Bitrate starts to matter only at the bottom. Below roughly 32 kbps, lossy encoders smear consonants badly enough that the damage survives resampling, which is why a heavily compressed voice note transcribes worse than a clean one. Between 64 and 320 kbps the differences disappear.

Two things do still matter, and neither is bitrate: how many times the file has been re-encoded, and the recording itself — microphone distance, room, overlapping speakers.

Duration, comfortably. At 128 kbps, 5 GB would hold about 87 hours, so the 8-hour cap arrives first by an order of magnitude. The largest MP3 processed here was 785 MB.

• Split by session, not by clock. Cut between recordings rather than mid-sentence — a word split across two uploads is lost to both.
• Variable bitrate is fine. VBR files upload and transcribe normally; a missing or wrong VBR header can make some players report the wrong duration, but it does not affect the transcript.