SoundStream is Google's pioneering end-to-end neural audio codec and the basis of Lyra's later versions. It introduced, in a fully learned codec, the residual vector quantization (RVQ) bottleneck that most neural codecs now use: a stack of vector quantizers that successively refine the encoded representation, letting one model serve a range of bitrates by using more or fewer quantization stages. Trained with adversarial and reconstruction losses, it reconstructs speech and music at very low bitrates with quality that traditional codecs can't match there. SoundStream's architecture became a template — EnCodec and others followed the same RVQ-autoencoder pattern — making it a foundational work in the move from hand-engineered to learned audio compression.
