Lyra and EnCodec Explained: The AI Codecs That Squeeze Sound Into a Few Kilobits

Lyra and EnCodec are AI audio codecs - instead of throwing away sounds you cannot hear, they learn to rebuild the sound from a tiny description.

Three small amber tiles casting a long shadow shaped like a microphone, titled Lyra and EnCodec Explained

Google built Lyra for voice calls on weak connections.

Meta built EnCodec for speech and music alike, and it now powers AI music tools.

Both squeeze audio into just a few kilobits per second - here is how they work, how they differ, and what you can actually use today.

How an AI codec is different from MP3

Classic codecs such as MP3, AAC and Opus study how human hearing works and drop the details you would not notice. That works well down to a point, and then the sound starts to crumble.

An AI codec uses two neural networks instead. The encoder turns the audio into a short list of numbers, called tokens, and the decoder has learned to rebuild convincing sound from them - even when very little data arrives.

The catch is that both sides need the trained model. Without the decoder, those tokens are just noise.

Lyra and EnCodec side by side

	Lyra	EnCodec
Made by	Google	Meta's AI research team
Built for	Speech on voice and video calls	Speech and music
Bitrates	3.2, 6 or 9.2 kbps	1.5 to 24 kbps
Sound	Mono, 16 kHz	Mono 24 kHz, or stereo 48 kHz for music
Licence	Open source, Apache 2.0	Open source, MIT
Use it if	You build a calling app for slow networks	You experiment with audio AI or extreme compression

Lyra: Google's codec for calls on bad signal

Google launched Lyra in 2021 to keep voice calls understandable when the connection is barely there. In 2022 it moved to a faster model based on Google's SoundStream, running at 3.2 kbps and up.

It handles speech only, and it is built for low latency on ordinary phones. The last release, 1.3.2, came out in December 2022, so the project is quiet rather than growing.

EnCodec: Meta's codec that handles music too

Meta released EnCodec in 2022. Its 24 kHz model works on any mono audio from 1.5 kbps, and a stereo 48 kHz model is trained on music alone.

Its biggest job today is not calls but AI. Meta's MusicGen, the text-to-music model, uses EnCodec tokens as the language it writes songs in.

It also makes for great stunts. In the Paper Tunes project, a maker used EnCodec to fit a two-minute song into eight QR codes printed on one sheet of paper.

Can you use them today?

Only if you are comfortable with code. Lyra is a developer library you build into an app, and EnCodec runs from Python - neither is something you install and double-click.

The practical pick: for small audio files you can play anywhere, use Opus Audio Codec. It keeps speech clear at very low bitrates, and in 2024 it gained optional AI features of its own for lost packets and low-bitrate speech.

Deciding between Opus and older formats? See how Opus compares with Ogg Vorbis.

Quick questions

Is EnCodec better than MP3?

At very low bitrates, yes - MP3 sounds poor below about 32 kbps, while EnCodec still sounds like music at 6 kbps. At normal bitrates MP3 and Opus sound more faithful, and they play on everything.

Is Lyra still used?

Google first used it in its Duo calling app, and the code is still public. It has had no new release since December 2022.

Are Lyra and EnCodec free?

Yes. Both are open source - Lyra under Apache 2.0 and EnCodec's code under the MIT licence.

Can VLC or my media player open these files?

No. There is no standard file format for either codec, so ordinary players cannot open them. You need the codec's own decoder.

Does Opus use AI now?

Partly. Opus 1.5 added optional neural features that hide lost packets and clean up speech at low bitrates, while staying compatible with every existing Opus player.

The bottom line

Lyra and EnCodec show where audio compression is heading - AI that rebuilds sound rather than trimming it. For files you want to play today, Opus is still the answer.