Best Audio Format for Transcription
MP3, WAV or M4A? The format matters less than you think: the model reads 16 kHz mono. The recording, re-compression and file size matter more.
The Scriptorium team ·
For transcription, the best audio format is the original recording, in whatever format it came, as long as it is not heavily compressed. MP3, WAV, M4A and most others work. The format hardly changes accuracy, because our engine converts every file to 16 kHz mono audio before the speech model sees it (SAMPLE_RATE = 16000 in the engine, and the faster-whisper decoder resamples to mono). What does change the result is the microphone and the room, and what changes the upload is file size.
Does WAV give a more accurate transcript than MP3?
Not because of the format. A WAV recorded at 48 kHz contains detail the model never receives, since it is resampled to 16 kHz first. A WAV does help in one way: it has not been through lossy compression, so a clean WAV is a safe copy to keep. If your recording is already an MP3 or M4A, converting it to WAV afterwards adds nothing; the loss has already happened.
What does change the result?
- Close microphone, quiet room. This outweighs any format choice. See how to improve transcription accuracy.
- Re-compressing a lossy file. Converting MP3 to MP3, or exporting a voice note from a chat app and sending it through another one, discards detail each time. Keep the original and upload that.
- Very low bitrates. We have not measured accuracy by bitrate, so we give no threshold. If you must shrink a file, make a copy and test a short clip first.
- Channels and sample rate in the original. A speaker who is only on the left channel, or a file recorded at a very low sample rate, is a problem that conversion cannot fix.
File size is the real difference
Size per minute follows from the format's data rate (uncompressed audio is sample rate x bits x channels; MP3 and AAC at a fixed bitrate are bitrate / 8). We checked the arithmetic by encoding 60-second test files with ffmpeg. The no-account preview takes files up to 25 MB and 15 minutes, so the table shows which limit you hit first.
| Format | Megabytes per minute | Fits in the 25 MB preview |
|---|---|---|
| WAV, 16 kHz, 16-bit, mono | 1.92 | about 13 minutes (size limit) |
| WAV, 48 kHz, 16-bit, stereo | 11.52 | about 2 minutes (size limit) |
| MP3 or AAC at 128 kbps | 0.96 | 15 minutes |
| MP3 or AAC at 64 kbps | 0.48 | 15 minutes |
Variable-bitrate files (many M4A voice memos) vary around their average, so these are guides. A WAV from a phone or recorder is often the biggest file you have; if it is over the limit, a 128 kbps MP3 is a sensible copy to upload, and you keep the WAV. The per-format pages on the formats hub list what the site accepts, and the hour-long guide has the ffmpeg commands to shrink or split a file.
A quick way to choose
- Do you still have the original recording? Upload that.
- Is it over 25 MB for the preview? Make a 128 kbps MP3 copy for the upload and keep the original.
- Is it a video? Extract the audio first: it is far smaller, and the picture is not used for transcription.
- Is it a chat-app voice note? Save the original file from the app rather than a screen recording of it.
ffmpeg -i original.wav -ac 1 -b:a 128k copy.mp3That command mixes to one channel and writes a 128 kbps MP3. Check that the copy plays and that the voices are still clear before you upload it.
Checked on 2026-10-26. The engine's sample rate and the faster-whisper decoder (both read in the code, mono at 16 kHz), our upload limits (from the same module as the site's other pages), and size arithmetic that we confirmed with ffmpeg test files. No outside page was needed for these claims. Not measured: how accuracy varies with bitrate. Menus, prices and limits change; confirm on the vendor's page before you rely on them.