Audio Tokenization

Audio tokenization turns sound into sequences of discrete codes or continuous latents that models can compress, predict, retrieve, or generate. Common approaches use neural codecs with vector quantization, self-supervised speech features, syllable- or text-aligned units, and learned continuous bottlenecks; some combine semantic and acoustic streams. These representations are tested on reconstruction, speech recognition and synthesis, audio understanding, music generation, and streaming tasks. A central tension is how to make tokens compact and predictable without losing acoustic fidelity, linguistic content, speaker information, or timing, and the best granularity often depends on the task.