Convert MOV to OPUS

Recorded interviews accumulate faster than anything else in a research project. A day of fieldwork is four or five cards, a documentary is hundreds of hours, an oral history programme collects them for years. They come off cameras as MOV files, and after the transcript exists almost nobody ever looks at the picture again.

What has to survive is the sound: for checking a quote, for re-transcribing when the first pass was wrong, for the archive that will outlast the project. Keeping the video for that is enormously expensive — a single 4K hour is comfortably ten gigabytes — and keeping WAV is not much better. Opus is the format that makes the whole collection something you can actually store, back up and search.

Opus is not a general audio codec with a low-bitrate mode bolted on. It was standardised as RFC 6716 by the IETF for real-time communication, and it contains two engines: a speech model derived from SILK and a music model derived from CELT, switching between them according to what it hears.

The practical result is that a single speaker at 38 kilobits a second is intelligible and pleasant, which no other format on this site manages. MP3 at that rate is unusable. That is the difference between an archive you can keep on one drive and an archive that needs a shelf, and it is the entire technical argument for this page.

Roughly 38, 64 and 128 kilobits a second, which is about 17 MB, 29 MB or 58 MB an hour. A hundred hours of fieldwork therefore fits in about 1.7, 2.9 or 5.8 gigabytes — small enough to hold on a laptop, back up twice and send to a collaborator over an ordinary connection.

Balanced is the default and the right answer for interviews recorded on a camera in an uncontrolled room. Drop to the small band only for a clean single voice recorded close, and go to the high band when there is music, ambience or several people talking across each other, where the extra bits genuinely buy separation.

Nothing here improves on what was recorded. Camera audio is a microphone a few centimetres from the lens, picking up the room, the air handling and the operator, and the codec faithfully preserves all of that. Where an interview was recorded on a separate recorder, use those files instead — the camera track is the sync reference, not the recording.

It is also a second compression whenever the camera wrote AAC, which phones always do and many cameras do. On speech that is inaudible; the point of mentioning it is that no bitrate here recovers a badly placed microphone, and choosing the high band on a poor recording spends storage on preserving the air conditioning very accurately.

Automatic transcription handles Opus without complaint in most services, and the small file size makes the upload to whichever service you use a fraction of what a video would be. Where a service refuses the format, convert one copy to WAV for that step rather than compromising the archive format.

Two things help the accuracy more than the bitrate does. Convert each interview as one file rather than splitting it, because a model uses the surrounding speech to disambiguate; and keep the filename identical to the video, so a timecode in the transcript can be traced back to the footage it came from when somebody eventually wants a clip.

Every audio track the browser can decode is written, in the order the camera recorded them, so a body running a radio microphone on a second track keeps both. Not every player offers a track menu for a multi-stream Ogg, so convert one file from the shoot and open it in whatever the project actually uses before committing to the other thirty-nine.

Everything outside the audio is discarded: the picture, the timecode track a professional body writes, the location, any subtitles. A handful of descriptive fields do survive, as the next section describes, but for a research archive this is still the point at which the metadata has to move somewhere else — a spreadsheet, a database, or the filenames themselves — and doing that at conversion time is far easier than reconstructing it later.

Title, description, artist and creation date are read out of the MOV’s metadata atoms and written into the Opus file as Vorbis comments. What a camera actually fills in is the creation date and rarely anything else, so the practical position is unchanged. A folder of Opus files with camera-generated names is an archive nobody can use, and the fix is a convention applied before the conversion rather than after: date, location, speaker, part number, in that order, so that a directory listing sorts itself into something meaningful.

If the project will last, fill in the remaining Vorbis comments with a tagging tool. Files get separated from the spreadsheets that describe them, and a recording that says inside itself who is speaking and when it was made is the one that survives being copied between three institutions over twenty years.

The limit is on the video handed over rather than the audio produced, and camera video is enormous — around 170 MB a minute at 4K. In practice this converts individual takes and short interviews, not a two-hour session recorded in one file.

For a project-scale job the extraction belongs where the footage already is: an editor or a batch transcoder can pull audio from a whole folder in one pass, and can select the track while it does. Use this page for the files on the machine in front of you and for checking that the settings you plan to run over everything actually give you what you want.

Reading the MOV, decoding it and encoding the Opus all happen in this browser tab. Nothing carrying the recording leaves the machine, and the network tab shows that during a conversion.

For this material it is a compliance question rather than a preference. An interview recorded under a consent form usually names who may hold the recording; an ethics approval names where it may be stored; a journalist’s source expects the recording not to be handed to a third party at all. A free converter that uploads the file breaks all three quietly, and nobody finds out.