codec models

Neural audio codecs compress audio into discrete tokens, enabling efficient storage and language model-style generation. **How it works**: Encoder compresses audio waveform to low-bitrate discrete codes, decoder reconstructs from codes. Vector quantization creates codebook of audio tokens. **Key models**: EnCodec (Meta), SoundStream (Google), DAC (Descript Audio Codec). **Technical details**: Residual Vector Quantization (RVQ) uses multiple codebooks for refinement, convolutional encoder/decoder, trainable codebooks. **Compression rates**: 1.5-24 kbps (vs 1400 kbps for CD), extreme compression with good quality. **For generation**: Audio tokens become vocabulary for language models. Generate token sequences, decode to audio. Foundation for AudioLM, MusicLM, Bark. **Advantages**: Unified representation for all audio (speech, music, sounds), compatible with transformer architectures, efficient generation. **Applications**: Audio compression, audio generation, neural voice synthesis, music generation. **Comparison to traditional codecs**: MP3/AAC use hand-designed transforms, neural codecs learn optimal compression. Revolutionary for audio AI.