AudioLM is the paper that convinced a lot of people audio generation should look more like language modeling than signal processing. Google's framework maps raw waveforms to discrete tokens, then trains a language model over them, casting audio continuation as next-token prediction. The clever bit is the hybrid tokenization: semantic tokens from a masked audio language model handle long-term structure, while acoustic tokens from a neural codec deliver high-fidelity synthesis. Together they solve the tradeoff that stumped earlier approaches.

Trained on raw waveforms with no transcripts or labels, AudioLM generates speech continuations that stay syntactically and semantically coherent while preserving speaker identity and prosody for unseen voices. It also generalizes beyond speech, producing plausible piano continuations without any symbolic music representation. This is the direct intellectual ancestor of VALL-E, MusicLM, and a wave of token-based TTS systems, so understanding the two-stage tokenizer design pays dividends when reading anything that came after. If you are building or evaluating token-based generative audio, this review is a worthwhile primer on the ideas that reshaped the field.