Neural codecs quietly became the foundation layer of modern generative audio, and SoundStream is one of the papers that got the ball rolling. Google's end-to-end design pairs a fully convolutional encoder/decoder with a residual vector quantizer, all trained jointly with adversarial and reconstruction losses borrowed from the TTS and speech-enhancement worlds. The headline result is striking: at 3 kbps, SoundStream beats Opus at 12 kbps and closes in on EVS at 9.6 kbps on 24 kHz audio, across speech, music, and general sound.

Beyond the compression numbers, the design choices matter for anyone shipping voice tech. Structured dropout across quantizer layers lets a single model span 3-18 kbps without retraining, and the architecture runs in real time on a smartphone CPU with streamable inference. The paper also folds background noise suppression directly into the codec at either the encoder or decoder side, with no added latency. If you are working on neural TTS tokens, low-bitrate calling, or on-device audio pipelines, this walkthrough is a solid grounding in the ideas that later powered AudioLM and its descendants. Hit play for the architecture breakdown and the subjective evaluation results.