Convert OPUS to FLAC

FLAC is lossless in the strict sense: decoding it returns the identical samples that went in, not an approximation. That property is exact and it is a statement about the process, not about the recording. What went in here is the output of an Opus decoder — the audio that survived the original encoding, with everything Opus removed still absent.

So the result is a large file whose contents are precisely the contents of the small one. Nothing is repaired, and no later step can tell the difference between this FLAC and one made from a real master except by looking at the audio itself. If a better source exists anywhere — the session, the original recording, the file the Opus was made from — that is what belongs in the ingest, and finding it is worth more than any conversion.

A FLAC stream header carries an MD5 of its decoded audio, computed at encode time. It is the single field that makes FLAC an archive format rather than a smaller copy: years later a decoder can reconstruct the samples, hash them, and prove that a drive has not silently corrupted a block. That is a genuinely valuable property and it works exactly as advertised here.

What it cannot do is reach back past the conversion. The hash covers the samples this encoder was handed, so it certifies a faithful copy of a lossy decode. A collection of transcoded FLACs will all verify cleanly forever while containing audio several generations from the source. Where provenance matters, record in the metadata that the file was derived from Opus, because the checksum will not say so and the file otherwise looks like a master.

Opus works on a fixed 48,000 Hz clock. A decoder reports 48 kHz regardless of what the encoder was originally handed, so a voice note recorded at 16 kHz and a music file encoded from a 44,100 Hz source both arrive here as 48 kHz audio, and the FLAC records that rate.

This is the detail that most often gets a deposit rejected. Ingest specifications and dataset schemas frequently name a rate — 44,100 Hz for a music corpus, 16,000 Hz for a speech one — and the converted file will not match it. Resample in an editor after converting, and note that downsampling a 48 kHz decode to 16 kHz is a legitimate step while upsampling adds nothing but bytes.

There are two cases and both are about requirements rather than fidelity. The first is a hard whitelist: an ingest, a repository or an analysis service that accepts FLAC and WAV and will not take Opus, usually because its pipeline was built around decoders it controls. Complying is the whole job and there is no argument to be had.

The second is a processing chain. If the audio is going to be filtered, normalised, segmented and written out several times, doing that work on a lossless intermediate stops each step compounding a fresh generation of lossy damage. Convert once, process in FLAC, and encode once at the end. What is not a reason is storage: keeping the Opus is smaller, and keeping both is ten times the disk for no additional information.

An hour of 64 kbps Opus is about 29 MB. The same hour as FLAC is roughly 200 to 350 MB, which is a wider range than usual because FLAC compression is a property of the material rather than a setting: sparse, quiet, predictable audio compresses hard, and dense loud material barely moves.

Decoded lossy audio sits at the easier end more often than you would expect. A lossy encoder has already removed low-level detail and high-frequency content, which is exactly the material that defeats FLAC's predictor, so a transcode frequently compresses better than a genuine recording of the same music. That is not a benefit — it is a measurable trace of what the Opus encoder took out.

The registry records Opus as supporting up to 255 channels and FLAC as specified to a maximum of eight. In this direction that ceiling is real rather than theoretical: mono, stereo, quadraphonic and 7.1 all fit comfortably, and anything wider has nowhere to go.

It bites on ambisonic and immersive material rather than on anything a person recorded on a phone, and Opus is one of the few codecs people actually use for wide channel counts because it handles them well. If a file is a first-order ambisonic capture or an object-based mix, check the channel count before converting a set — a target that cannot hold the layout is a different problem from one that degrades it, and it will stop rather than fold anything down.

Metadata usually loses something at a format boundary and here it does not. Opus is carried in an Ogg container that holds Vorbis comments — a plain list of uppercase keys with text values — and FLAC uses the same structure in its own tag block. Title, artist, date, description, licence and any custom field arrive spelled exactly as they were.

Embedded artwork crosses as well, in a different shape at each end: the Ogg holds the cover as a METADATA_BLOCK_PICTURE comment, and the FLAC gets a native PICTURE block written ahead of the tag block, which is where RFC 9639 puts it. For an ingest that is more useful than it sounds: descriptive metadata written into the source file survives into the deposit without being re-entered, which removes the transcription errors that a manual re-keying step reliably introduces.

The instinct after this conversion is to delete the small file, and it is the wrong one. The Opus is the earlier generation. It contains the same audio in a tenth of the space, and any future conversion — to a different lossless format, to a different rate, to whatever the next system wants — is better made from it than from a decode of it.

The FLAC is a deliverable produced for one requirement. Treat it as output: regenerable in a minute, safe to delete once the ingest has accepted it. The only situation in which the FLAC becomes the thing worth keeping is one where it has been edited, and at that point it is a new recording rather than a copy of an old one.

Some browsers expose a FLAC encoder through WebCodecs and use it directly. Where none exists, a WebAssembly FLAC encoder is fetched the first time you convert to FLAC, once, and only on the pages that need it. Decoding Opus needs nothing at all, because every browser implemented it for calls.

The audio itself never crosses the network in either direction. Files are capped at 100 MB each — which for Opus is several hours of speech, though the FLAC coming out of it will be far larger in memory — and a hundred can be dropped at once and taken back as a ZIP. For research recordings held under a consent agreement, that the conversion is local is the reason it can happen here at all.