Convert WebM to AAC

There are two things people mean by "AAC". One is the codec, which is also what sits inside every M4A, every MP4 and every streamed track from a major service. The other is the elementary stream: the encoded frames on their own, with no container around them. This conversion produces the second.

That distinction decides whether you are on the right page. If the goal is a recording you can add to a music library, give a title and press play on, the M4A conversion is the one you want and this one will make your life harder. If something in a pipeline has told you it wants AAC and rejected an MP4, this is the file it was asking for.

ADTS stands for Audio Data Transport Stream, and its design goal is that any point in the file is a valid starting point. Every frame carries a seven-byte header stating the profile, the sample rate and the channel configuration, so a decoder that joins halfway through knows immediately how to interpret what follows.

That is exactly what broadcast and streaming need. An HLS segmenter can cut a stream at a frame boundary and each piece stands alone. Two streams can be concatenated with nothing more than a byte-level join and the result is valid. An embedded player with a few kilobytes of buffer can start decoding without parsing an index it does not have room for. None of that is true of an MP4, where the index lives in one place and describes everything.

Everything the container was doing is now absent. There is no stored duration, so players compute the length by scanning frames or, more often, estimate it from the bitrate and get it wrong on variable-rate audio. There is no seek table, so jumping to a position is a calculation rather than a lookup. There is nowhere to write a title, an artist or a date.

For a machine reading the stream once, from the beginning, none of that matters — which is why the format survives. For a person with a media player it matters immediately, and it is the reason an `.aac` file so often shows a plausible but wrong duration in a file browser. If that bothers you, the file wanted to be an M4A.

AAC was standardised by MPEG in 1997 as ISO/IEC 13818-7 and has been the default for delivery ever since — broadcast, streaming, mobile, Apple’s catalogue. It handles up to 48 channels, though anything from a browser recording is mono or stereo.

What went in was Opus, which is a newer and generally better codec that almost nothing outside a browser accepts. Converting between them is a full decode and re-encode: two lossy codecs cannot exchange data directly, so this is a second generation of compression. For speech it is inaudible, and for a pipeline it is usually irrelevant next to whatever compression the delivery chain applies afterwards.

The common shape of this problem is a web application that records through the browser. `MediaRecorder` writes WebM with Opus because that is what browsers are required to support, and the recording then has to enter a backend that was designed around AAC — an ingest service, a transcoding farm, a telephony platform, an app bundle with a fixed asset format.

Converting on the client side, before the upload, is often the cheaper architecture. The file that crosses the network is already in the format the backend wanted, no server has to run a transcoder per recording, and there is nothing to queue. That is the same reasoning that put this conversion in the browser in the first place.

Opus operates at 48 kHz, so a browser recording is 48 kHz and the AAC written from it is 48 kHz. The channel count is likewise whatever was captured — mono for a microphone, stereo for anything that recorded system audio.

Nothing here resamples, and that is deliberate: resampling is a signal-processing decision with its own quality trade-offs, and a converter that silently did it would be hiding a step. If a target platform requires 44.1 kHz, or a specific channel layout, do that conversion in the tool that has the requirement, where the resampler is one you chose.

Any video track in the WebM is discarded outright rather than ignored, which is the point when the source is a screen capture with narration. Only the first audio track survives; a recording that captured a microphone and system audio as separate tracks yields one of them with no way to choose.

For a pipeline this is worth asserting rather than assuming. If the recorder on the other end of your system writes two tracks, verify which one arrives before the behaviour becomes a bug report about missing audio three weeks later.

It is worth ruling out the wrong use before it costs somebody an afternoon. Pointing an `<audio>` element at an `.aac` file is unreliable: browsers vary in whether they will parse a raw ADTS stream, and the ones that do often report a wrong duration because there is nothing in the file to tell them the real one. The seek bar then behaves oddly and the bug looks like a player problem.

For anything played in a browser, the same audio in an M4A works everywhere and behaves correctly, because the container carries the duration and the index. Keep the ADTS stream for the pipeline that asked for it and use a container for anything a person interacts with — that division is what the two formats are actually for.

Up to a hundred WebM files convert in one pass, each with its own progress, downloading together as a ZIP. Because there is no server involved, there is no daily quota and no queue; the per-file ceiling is 100 MB.

That is the practical difference for anyone doing this as part of a job rather than once. A free converter that uploads will cap you somewhere between five and twenty files a day; here the hundred can be dropped again immediately, and it behaves the same on the hundredth file as on the first.