Convert OPUS to AAC

The file this page writes is Audio Data Transport Stream framing: AAC frames one after another, each carrying a short header that repeats the sample rate, the channel configuration and the profile, with nothing wrapped around them. No index, no duration field, no metadata area. The other file called AAC is an M4A — the same encoded audio inside an MP4 container that carries all three.

Which one you need is not a matter of taste and cannot be inferred from the word. Documentation that names ADTS, an elementary stream, or a `.aac` extension means this page. Documentation that says "AAC audio" and expects a length, a title or a seek bar means the M4A. Sending the wrong one is the single most common failure in this conversion, and it usually surfaces as a player reporting the wrong duration rather than as an outright rejection.

ADTS exists so that a stream can begin at any frame. That property is what segmenters, playout automation, announcement and IVR players and embedded parsers are built around: they cut audio into pieces, start playing partway through, and hand fragments to a decoder that has no container to consult. A header on every frame is what makes that safe.

The same design supports concatenation. Two ADTS files whose sample rate and channel count match can be joined byte-wise and the result plays as one continuous file, with no re-encoding and therefore no additional loss. Prompt systems assemble sentences from fragments exactly this way. It is the strongest reason to prefer a bare stream when you have the choice, and it is worth checking that your fragments really do share parameters before relying on it.

Opus runs on a fixed 48,000 Hz clock. A decoder reports that rate whatever the file was encoded from, so the AAC produced here is 48 kHz even when the source was a 16 kHz voice recording. Nothing on this page resamples, and that is deliberate — resampling is a decision with audible consequences and belongs in a tool where it can be heard.

It matters most in the direction this pair is often used. Telephony, IVR and announcement systems commonly specify 8,000 or 16,000 Hz, because that is what the transport carries, and a 48 kHz file will either be refused or resampled on ingest. Do it yourself in an editor or a batch tool after converting, so the result is checked rather than discovered when a prompt sounds wrong on a live line.

Opus turns up in this position for three reasons. Messaging platforms and voice bots record in it, because it was designed for real-time speech and is what WebRTC mandates. Download tools produce it, because it is the best stream available without re-encoding. And synthesis and text-to-speech services increasingly emit it, for the same bandwidth reasons.

The source bitrate sets the ceiling on everything downstream. A prompt synthesised at 64 kbps or better converts to a 192 kbps AAC that no listener will fault. A voice note captured at 24 kbps for a phone call is already at the bottom of the codec's range, and no band on this page improves it — the highest setting only stops the second encoder adding to the damage.

The Quality control resolves to a real bitrate inside the encoder, and AAC is offered a fixed set: Small file writes 96 kbps, Balanced 128, High quality 192. Nothing above 192 is available on this path, which is a limitation of the encoder rather than of the codec, and worth knowing if a specification names 256.

Requirements written for AAC are usually ceilings imposed by a transport rather than quality targets — 64 kbps because a channel is narrow, 128 because a budget was set years ago. Pick the band at or just under the figure named. Where nothing is specified, High quality is the right default for a second-generation encode, because the point of the headroom is that the encoder then contributes almost nothing of its own.

Speech recorded for a system is normally mono at source, and it should stay that way: two identical channels convey the same information at twice the bitrate, and a prompt player that expects mono may reject or fold a stereo file unpredictably. Nothing here promotes a mono file to stereo, so a mono Opus produces a mono AAC.

The ADTS header records the channel configuration on every frame, which is what makes the concatenation trick above safe — a mono fragment cannot accidentally be joined to a stereo one without the mismatch showing. If a batch contains both, that inconsistency is worth resolving before conversion rather than after, because it is much easier to see in a source directory than in a stream of frames.

There is no duration anywhere in an ADTS file. Software that needs one estimates it, either by scanning every frame or by dividing the file size by an assumed bitrate, and variable-rate material defeats the arithmetic. That is why players sometimes report the wrong length for a `.aac` file or refuse to let you seek in it.

For a prompt system this is irrelevant, because the platform measures the audio when it ingests it and stores the figure in its own database. For anything where a person sees a progress bar it is a visible defect, and it is the clearest practical sign that a bare stream was the wrong deliverable. The M4A carries the duration in its header and does not have this problem.

Opus is carried in an Ogg container holding Vorbis comments — an open list of uppercase keys with text values — and a synthesised prompt or an exported recording quite often has a useful description in there. ADTS has no metadata area at all, so every field is dropped and the result carries only sound.

In this context that is usually correct: prompts are identified by path and by the platform's own record, not by embedded tags. Where the description matters — a licence, an attribution, a version number for a prompt set — copy it into the file name or a manifest before converting, or use the M4A target, which has atoms for all of it and holds identical audio.

Files are capped at 100 MB each — for Opus that is hours of speech — and a hundred can be dropped at once and taken back as a ZIP, which fits how prompt sets and bot recordings actually arrive. The work is processor time on your own machine rather than bandwidth, so a large batch is worth starting and leaving.

Opus decodes natively in every browser because WebRTC made it mandatory. Chrome, Edge and Safari also encode AAC natively; Firefox does not, so a WebAssembly encoder is fetched once there. That package is the only thing that crosses the network — recorded calls, customer prompts and anything else in the batch stay on the machine, which you can confirm in developer tools while a conversion runs.