Convert MKV to OPUS

A meeting or interview captured to Matroska is mostly picture, and the picture is the least interesting part of it — a static shot of a table, a shared screen that changes twice an hour, a grid of thumbnails. It compresses badly precisely because nothing moves in an interesting way, and it is why a two-hour recording is measured in gigabytes.

Discarding it is not a compromise, it is the whole conversion. Nothing about the sound is decided by how long or how large the video was; what comes out is an audio file whose size depends on the duration and the quality band and nothing else. That is why the reduction here is startling in a way that compressing the video would never be.

Opus was standardised in 2012 as RFC 6716 and designed around real-time speech, which is a different problem from music: it has to stay intelligible when the bitrate is low and the connection is imperfect. Every browser shipped an encoder for it because WebRTC needed one, which is also why this conversion starts immediately where an MP3 has to download an encoder first.

For a transcription job that design matters. Where MP3 at a low bitrate blurs consonants into each other, Opus keeps the parts of speech a recogniser is actually keying on. The practical consequence is that you can choose the small band without a real accuracy cost, which is exactly the choice an upload ceiling forces.

Recognition errors come from the recording, not the compression. A microphone across a table, a hard-surfaced room with a second of reverb, two people talking over each other, an air conditioner — those are what produce a transcript full of guesses, and every one of them was baked in before the file reached a converter.

It is worth saying because the instinct when a transcript is bad is to convert again at a higher quality band, and that will not help. If the recording is poor, the levers that exist are on the audio itself: convert to a WAV, clean it in Audacity or iZotope, and send that. If it is good, the small band will transcribe as well as the high one and upload in a third of the time.

The Ogg container this writes into has no one-track limit, so every audio track the browser can decode goes into the file. That is unusual — an MP3, an AAC, a WAV or a FLAC would keep one — and for this reader it is a trap rather than a benefit, because a recognition endpoint will read the first stream and ignore everything after it.

On a single-microphone recording there is nothing to think about. On a conference capture with a lectern feed, a room feed and a translation channel, the first stream in the file may not be the one with the speaker on it. Matroska’s own default-track flag is not consulted here, so check the order with MediaInfo, and extract the track you want with MKVToolNix before converting if it is not first.

Three bands: small file, balanced, high quality, defaulting to balanced. The control is a band rather than a bitrate because the number that means "good" is codec-specific, and Opus reaches at a given size what MP3 needs considerably more to match.

For speech that is going to a recogniser rather than to a listener, start at small and only move up if the transcript disappoints. For an interview you also intend to publish as audio, balanced is the sensible default and the difference in size against small is not large enough to argue about. The high band exists for material with music or performance in it and is rarely the right answer for a meeting.

Both targets on this site write Opus audio into an Ogg container using the same muxer, so the output is byte-for-byte the same and only the filename differs. Nothing is gained or lost by picking one over the other in terms of what the file contains.

The name still matters, because a good deal of software decides what a file is before it looks inside it. Upload forms, ingest scripts and a surprising number of APIs branch on the extension, so use whichever one the service’s documentation lists. If it lists neither, .ogg is the more widely recognised of the two.

A conversion that stops with a message saying nothing in the file could be converted has run into a soundtrack the browser cannot decode — AC-3, E-AC-3, DTS or Dolby TrueHD, none of which browsers ship decoders for. With the picture already discarded, there is nothing else to write.

Recorders do not produce those. A meeting captured by OBS, Zoom, a conference system or a phone is AAC or Opus and converts without incident. It is disc-sourced Matroska that fails, which for this reader means an archival recording of an older event on a DVD rather than anything current — and those need their audio converting once with a desktop tool before this page can help.

The point of this conversion is usually to hand the result to a service. That makes it worth noticing what the alternative route costs: uploading a gigabyte of recording to a converter, waiting, downloading the audio, and then uploading that to the recogniser — the same confidential meeting passing through two third parties instead of one.

Here the decode and the Opus encode happen in the browser tab on your own processor, so nothing carrying the recording moves until you choose to send the finished audio. The free tier takes an MKV up to 100 MB, which fits a short meeting and not a long one; for anything larger, extracting the audio locally with ffmpeg and converting the result here is the sequence that keeps both properties.