EnCodec

EnCodec is Meta AI's open-source neural audio codec, introduced October 2022 in the paper 'High Fidelity Neural Audio Compression'. It uses a streaming encoder–residual-vector-quantization–decoder architecture trained end-to-end with multi-scale STFT discriminators as perceptual loss, compressing audio to roughly one-tenth the size of MP3 at comparable quality — 64kbps output that holds up against MP3 at far higher bitrates. It was the first neural codec demonstrated on 48kHz stereo audio.

Beyond compression, EnCodec became foundational infrastructure: its discrete audio tokens power generative audio models (notably Meta's own MusicGen and AudioGen pipelines) and it remains a standard tokenizer reference in audio LM research. The honest picture: as a pure codec it has been surpassed by newer options like Descript's DAC on some quality metrics, and the pre-trained checkpoints target 24kHz mono or 48kHz stereo at fixed bandwidths — but for an MIT-licensed, CPU-realtime codec with a huge ecosystem, it is still the default starting point.