Which audio and video formats can be transcribed
Updated
The format does not change the transcript. Everything is decoded by the browser's own decoders and resampled to 16kHz mono before the model hears it, so a 96kbps MP3 and a studio WAV of the same recording produce the same text. What matters is whether the browser can decode the file at all. MP3, M4A and AAC, WAV, FLAC, OGG and Opus work, and so does the soundtrack of an MP4, MOV, MKV or WebM. AMR and 3GP from older phones, WMA, AVI, and the proprietary formats dictation recorders write do not, and need converting with ffmpeg or VLC first.
Container and codec are two different questions
A media file is a container holding one or more encoded streams. The extension names the container: .mp4, .mkv, .ogg, .wav. The audio inside is a codec: AAC, Opus, MP3, PCM. Both have to be understood, by different pieces of software, and a file fails on either.
Reading the container is done by mediabunny, the library this site uses for all media work. It parses MP4 and M4A, QuickTime .mov, WebM and Matroska .mkv, MP3, WAV, Ogg, FLAC, raw AAC in .aac, and MPEG transport streams in .ts. It does not parse AVI, AIFF or the ASF container that WMA and WMV use, and a file in one of those is refused at the container step before any codec is looked at.
Decoding the stream is then handed to the browser through WebCodecs, so the codec list is whatever your browser ships. In Chrome and Edge, and in Firefox since it added WebCodecs, that is AAC, MP3, Opus, Vorbis, FLAC and uncompressed PCM. Safari has had WebCodecs since 16.4, but this site has not been verified there and does not claim it.
What works, by the file you are likely to have
The short version is that anything made in the last ten years by a phone, a meeting app or a podcast workflow works.
- MP3. Podcasts, exported recordings, anything old. Decodes everywhere, and the bitrate makes no difference to the transcript
- M4A and AAC. iPhone voice memos, Android recorders, Zoom's audio_only file, and the soundtrack of nearly every MP4
- WAV. Uncompressed, so the files are large, but size is not a limit here. A two-hour 48kHz stereo WAV is about 1.3GB and is fine. The limit is length, at 2 hours per recording, and it is about memory rather than bytes on disk
- FLAC. Lossless archives and some field recorders. Decodes directly
- OGG and Opus. WhatsApp and Telegram voice notes, Discord recordings, and most things made by open-source tools. Opus was designed for speech and a 24kbps voice note transcribes as well as a WAV would
- MP4, MOV, MKV, WebM and TS video. Only the soundtrack is read for a transcript, so a phone video, a screen recording from a browser and a lecture capture all work the same way as audio does
What does not, and the message you will see
A file that cannot be decoded is refused with a sentence naming the codec that was found, rather than with a spinner that never ends. The suggestion it makes is to convert to MP4 with AAC audio, or WebM with Opus, both of which every browser decodes.
The formats that reach that message are mostly old or proprietary.
- AMR, usually inside a .amr or .3gp from an older phone's call or voice recorder. No browser decodes it
- WMA and WMV, from Windows-era recorders and Movie Maker. The container is not parsed and the codec would not decode anyway
- AVI and AIFF. The container is not parsed, whatever is inside
- AC-3 and E-AC-3 surround soundtracks, common in a file ripped from a disc. Chrome on Windows and Mac has no decoder for them, so the video opens and the audio does not
- ALAC, which iPhone Voice Memos writes when Lossless is switched on. Support is uneven, so treat it as a convert-first format
- DSS and DS2 from Olympus and Philips dictation recorders, which only their own software reads
The fix, and why it does not change the words
Conversion is one command with ffmpeg, which is free and runs on your own machine, so nothing is uploaded in the process either: ffmpeg -i input.amr -c:a aac output.m4a. VLC does the same from its Convert menu, and most recorders' desktop software exports MP3 or WAV.
Converting between formats will not make a transcript better or worse, and neither will choosing a lossless one to begin with. Whisper takes 16kHz mono audio and turns it into a spectrogram before it does anything else, so everything above 8kHz is discarded and stereo is folded to one channel whatever you hand it. A 64kbps AAC or Opus file keeps everything the model will use.
What does change the words is the original capture: the distance to the microphone, the room, two people talking at once, and the 8kHz telephone band that a call recording was squeezed through before it was ever saved. None of that comes back by converting, and the place to spend effort is the recording, not the file.