AudioLM
Google's framework for generating realistic speech and audio continuations
AudioLM is a research framework that Google published in September 2022, treating audio generation as a language modeling problem over discrete audio tokens instead of the usual spectrogram or waveform regression approaches. It combines two representations: semantic tokens from w2v-BERT, a self-supervised speech model, that capture long-range structure, and acoustic tokens from the SoundStream neural codec that preserve fine audio detail for high-quality synthesis. Trained only on raw audio with no text transcripts or annotations, it can continue a few seconds of spoken audio or piano music in a way that stays coherent over tens of seconds, something earlier audio generation methods struggled with. Google never released it as a product; it stayed a research paper and demo page, and Google DeepMind used its token-based approach as a building block for later work like MusicLM and parts of Lyria.