All Tools
E
OtherFreeOpen Source
ENCODEC
Neural audio compression at ultra-low bitrates
MIT
ABOUT
Traditional audio codecs like MP3 and Opus use hand-engineered psychoacoustic models that degrade rapidly at very low bitrates and do not adapt to the content type. EnCodec uses a learned end-to-end neural network with a residual vector quantizer (RVQ) and a transformer-based language model to compress audio at bitrates from 1.5 kbps to 24 kbps. It maintains perceptual audio quality far below the bitrates where traditional codecs break down, supports both mono and stereo streaming, and runs in real time on consumer hardware — enabling applications like high-quality music streaming over limited bandwidth and efficient storage of large audio datasets.
INSTALL
pip install encodecINTEGRATION GUIDE
1. Compress high-quality audio for streaming at very low bitrates (as low as 1.5 kbps) without perceptible quality loss
2. Reduce storage requirements for large audio datasets used in machine learning training pipelines
3. Enable real-time audio compression for live streaming applications on consumer hardware
4. Serve as the acoustic compression backbone for generative audio models (MusicGen, AudioGen)
TAGS
audio-compressionneural-codecaudiodeep-learningpytorchspeech-synthesismusic-generationreal-time