Audio and video file formats for transcription: one reference table
Which audio and video formats a browser can transcribe, what each carries inside, what is discarded before the speech model hears it, and when converting is worth it.
Every format on this page can be dropped on the transcriber as it is. Before the speech model hears a sample, the browser decodes the file, mixes it to one channel and resamples it to 16 kHz, so the differences between formats mostly vanish at that point: a 320 kbps stereo MP3 and a 48 kbps mono one arrive in the same shape. What is left to care about is whether the browser can decode the container and codec at all, how much the encoder threw away, and how big the file is to move around. This page is the reference; the linked posts go deeper on the formats people ask about most.
The table
| Format | Container | Usual codec inside | Lossy? | Decodes in the transcriber? | Where it comes from | Deeper post |
|---|---|---|---|---|---|---|
| MP3 | MPEG audio | MPEG-1/2 Layer III | Yes | Yes | Voice recorders, podcast downloads, editor exports | Transcribe MP3 to text |
| WAV | RIFF | PCM (uncompressed); occasionally others | No (for PCM) | Yes for PCM | Studio recorders, Audacity, text-to-speech tools | Transcribe WAV to text |
| M4A | MP4 | AAC; ALAC on some Apple exports | Yes (AAC) / No (ALAC) | Yes for AAC; ALAC depends on the browser | iPhone Voice Memos, Zoom, iTunes | Transcribe M4A to text |
| AAC | Raw ADTS stream | AAC | Yes | Yes | Some phone recorders, radio rips | See the M4A post; same codec, thinner wrapper |
| FLAC | FLAC | FLAC | No | Yes | Archival copies, audiophile rips | None yet; lossless, so treat it like WAV |
| OGG | Ogg | Vorbis or Opus | Yes | Yes | Older web recorders, some Android apps | None yet; Opus is the codec discussed in the WebM post |
| WebM | Matroska-based | Opus (audio), VP8/VP9/AV1 (video) | Yes | Yes | Browser recorders, screen capture, Google Meet | Transcribe WebM to text |
| MP4 / MOV | MP4 / QuickTime | AAC audio; H.264 or HEVC video | Yes | Yes; video is ignored | Phones, cameras, Zoom, Teams, screen recorders | Get a transcript from an MP4 |
| MKV | Matroska | Anything | Depends | Yes for AAC, Opus, Vorbis, MP3, FLAC, PCM | Screen recorders such as OBS | See the MP4 post; the same rules apply |
| AVI | RIFF | MP3, PCM or AAC audio; AC-3 on some cameras | Depends | MP3, PCM and AAC yes; AC-3 and DTS refused | Dashcams, older cameras, DVD rips | None; convert to MP4 if refused |
| WMA / WMV, FLV, MPEG-PS / TS, RealMedia | Various | Proprietary or legacy | Yes | Refused with a message | Old Windows recorders, Flash-era video, DVDs | Convert to MP4 or MP3 in VLC or HandBrake |
“Decodes in the transcriber” means the browser’s WebCodecs decoder, or on browsers without it the Web Audio decoder, can read the codec. A codec the browser cannot decode produces a clear message naming it, and nothing is uploaded either way.
What happens to every format before the model hears it
The speech model reads audio at 16,000 samples per second, one channel. The paper behind OpenAI’s open-source Whisper model states that all audio is resampled to 16 kHz before it becomes the 80-channel spectrogram the model consumes. The transcriber does that resampling in the browser, so:
- Anything above 8 kHz in your recording is discarded. Consonant detail lives below that, so speech is fine; the loss is to the parts of the signal your ears use for “quality”.
- Stereo is averaged to mono. Two microphones on two channels become one stream; speaker separation, if you want it, has to happen before or after transcription.
- Bit depth and sample rate above 16 kHz buy nothing. A 96 kHz, 24-bit WAV and a 16 kHz, 16-bit WAV of the same speech give the same transcript.
That is why the format matters less than people expect. What still matters is how much the encoder removed on the way in.
Lossy versus lossless, in practice
A lossless file (WAV with PCM, FLAC, ALAC) is an exact copy of what the recorder captured. A lossy file (MP3, AAC, Opus, Vorbis) keeps what a listener is most likely to notice and drops the rest. At moderate bitrates the dropped material is above the 8 kHz the model never sees anyway, so a good MP3 transcribes as well as its WAV. At very low bitrates the encoder starts smearing the quiet, high consonants that separate “fifteen” from “fifty”, and that is where transcripts go wrong. We have not measured the exact point where accuracy falls off and will not invent one; the MP3 post explains what Apple’s podcast guidance implies about safe bitrates.
The one rule that always holds: do not re-encode. Every lossy pass discards more, and a second encoder spends bits describing the first one’s artefacts. Use the file the recorder wrote, not a copy that has been through an editor and exported again.
When converting is worth it
| Situation | Convert? | To what |
|---|---|---|
| The transcriber refuses the file (WMA, FLV, AC-3 in AVI) | Yes | MP4 with AAC audio, or MP3, in VLC or HandBrake |
| The file is a 4 GB video and you only need the words | Optional | Nothing; only the audio track is read, so the size costs you disk space, not time |
| A WAV of a two-hour meeting is 1.2 GB and has to be emailed first | Yes | MP3 at 64 to 96 kbps mono; same transcript, one twentieth of the size |
| An MP3 sounds muffled | No | Converting cannot restore what the encoder removed; find the original |
| A phone recording is M4A | No | It decodes as it is |
| A WebM from a browser recorder has no duration | No | It decodes; the progress bar just cannot show a percentage until the end |
Sizes, for planning
An hour of speech weighs roughly:
| Format | Typical setting | About |
|---|---|---|
| WAV, 44.1 kHz, 16-bit stereo | Studio default | 600 MB |
| WAV, 16 kHz, 16-bit mono | What the model actually uses | 115 MB |
| FLAC of the stereo WAV | Lossless | 300 to 400 MB |
| MP3, 128 kbps | Podcast download | 58 MB |
| MP3, 64 kbps mono | Voice recorder | 29 MB |
| AAC in M4A, 64 kbps | Voice Memos | 29 MB |
| Opus in WebM, 32 kbps | Browser recorder | 14 MB |
These are arithmetic (bitrate times seconds), not measurements, and real files vary with the encoder’s settings. None of them matters to the transcriber’s memory use, which stays at about a minute of decoded audio however long the file is, on the standard streaming path.
Sources, checked 25 September 2026
- https://arxiv.org/html/2212.04356 - Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision”: section 2.2, audio resampled to 16,000 Hz, 80-channel log-Mel spectrogram
- https://developer.mozilla.org/en-US/docs/Web/Media/Guides/Formats/Containers - MDN media container guide: MP4, MPEG audio, Ogg, WAV, WebM, QuickTime and what each may carry
- https://developer.mozilla.org/en-US/docs/Web/Media/Guides/Formats/Audio_codecs - MDN audio codec guide: MP3, AAC, FLAC, Opus, Vorbis, ALAC and PCM characteristics
- https://podcasters.apple.com/support/893-audio-requirements - Apple Podcasts audio requirements, used for the bitrate discussion in the MP3 post