SoundStream
Lab & Provider: Google Research
Overview & Technical Scope
Neural audio codec developed by Google Research (Zeghidour et al., 2021) utilizing fully convolutional encoder-decoder architecture with Residual Vector Quantization (RVQ) and adversarial STFT discriminators to discretize audio into multi-rate acoustic token streams.
Fully Convolutional Encoder-Decoder + Residual Vector Quantizer (RVQ) + STFT Discriminators
Evaluated in Zeghidour et al. (2021) across subjective MUSHRA audio quality listening tests and objective metrics (PESQ, ViSQOL) showing state-of-the-art compression efficiency between 3 kbps and 18 kbps.
Acoustic discretization and reconstruction codec. Tokenizes continuous audio waveforms into discrete representation codes; does not model sequence semantics, language structure, or inter-species translation.
Note: Computational acoustic pattern modeling and next-token sequence predictions must not be conflated with semantic translation of non-human vocalizations into human language.