Convert audio for transcription: the best format for Whisper
The best audio format for transcription is a modest MP3: 128 kbps, mono. Whisper and the commercial tools built on models like it downsample everything to low-rate mono internally before a single word gets recognised, so a lossless studio file buys you nothing except a slower upload. A 2 GB Zoom recording, a phone voice memo, a dictaphone file from 2006: convert any of them into a small MP3 and every upload form will take it without complaint.
That’s the whole answer. What follows is the recipe: the exact settings, the fastest route from each source format, and why confidentiality should decide where the conversion happens.
The settings, at a glance
| Setting | Recommendation | Why |
|---|---|---|
| Format | MP3 | The one format every transcription tool accepts, from Whisper to court-reporting suites |
| Bitrate | 128 kbps | Transparent for speech; about 1 MB per minute of audio |
| Channels | Mono | Speech gains nothing from stereo, and one channel halves what the bitrate has to cover |
| Sample rate | Auto | Models resample internally anyway; forcing a rate adds a step and changes nothing |
Two of these deserve a sentence. Mono first: interview and meeting audio carries the same voice in both channels, so keeping stereo doubles the data for zero benefit. In mp3fy the mono switch is one tap in the options row. And 128 kbps because that’s already past the point where recognition accuracy improves; the longer argument is in which MP3 bitrate you actually need. Nobody is going to listen to this file critically. A model will eat it and produce text.
One exception worth knowing. If your recorder captured each speaker on a separate channel and your workflow uses that separation to tell speakers apart, keep stereo. For everything else (one mic in a room, a phone call, a video call), mono is free accuracy-per-megabyte.
The arithmetic that makes this work: at 128 kbps, speech runs about 1 MB per minute, so an hour-long recording lands near 55 MB, comfortably inside the typical caps transcription services put on uploads. The conversion side has no size or duration limits at all; a full-day hearing is just a longer progress bar.
From whatever you have
Match the source to the route. Every path below ends at the same 128 kbps mono MP3.
- Zoom, Teams and OBS recordings: these are video files whose picture track is dead weight for transcription. Convert MP4 to MP3 and a 2 GB meeting becomes roughly 50 MB of audio. Same words, a fraction of the upload.
- Phone voice memos: iPhone Voice Memos and most Android recorders write M4A. They convert on-device in seconds, right in the browser.
- WhatsApp and Telegram voice notes: OPUS under the hood, a format plenty of upload forms have never met. OPUS to MP3 also runs entirely on-device.
- Old dictaphones and court recorders: WMA, AMR and stranger things from the 2000s. Browsers can’t decode most of these, so they ride mp3fy’s cloud path: native ffmpeg does the conversion and the upload is deleted the moment your MP3 is ready.
- A whole case folder: drop everything in at once. Batch conversion works through the lot and hands back one ZIP of consistent MP3s, ready to feed the transcription queue in order.
Keep confidential audio confidential
Consider what’s actually on these recordings. Depositions. Source interviews a journalist promised to protect. Medical dictation covered by privacy law. These are precisely the files that should never transit a random converter’s servers. And most free converter sites are exactly that: servers you know nothing about, run by someone you can’t name, with a retention policy you can’t read.
For every browser-decodable format (MP4, M4A, OPUS, WAV and the rest), mp3fy converts on your device and the audio never leaves your machine. The only service that ever sees it is the transcription tool you deliberately chose: the one whose terms you read, or at least the one your firm signed off on. The handful of legacy formats that genuinely need the cloud path are uploaded for the duration of the conversion and deleted immediately after.
That’s a shorter chain of custody than emailing the file to yourself.
Convert, glance at the duration to confirm nothing was truncated, and send it off. The model neither knows nor cares what the recording started life as.