SoundStorm 并行音频生成

SoundStorm receives a conditioning sequence and predicts acoustic tokens that a neural codec can decode into audio. The paper presents it as an acoustic generator within an AudioLM-style hierarchy, replacing the coarse and fine autoregressive acoustic stages. Semantic tokens supply high-level conditioning, while SoundStream tokens represent codec information about the waveform. This is different from directly producing a text transcript or generating each raw waveform sample in sequence. The model uses bidirectional attention and confidence-based parallel decoding inspired by masked generative methods. It samples candidates for masked positions, retains selected confident predictions according to its schedule, and revisits unfilled positions in subsequent iterations. Parallel prediction across positions is compatible with iterative refinement; it does not mean every level is solved in one unconditional step. Residual vector quantization provides a hierarchy. Coarse levels establish information that finer levels refine, with each quantizer representing a residual left by earlier ones. The stated decoding procedure advances from coarse to fine, proceeding to a subsequent level after earlier levels have been sampled. A short acoustic prompt can preserve speaker-voice information, while semantic conditioning provides a different control over content. For dialogue, the reported demonstration combines speaker-turn transcripts and speaker voice prompts through the surrounding system. The abstract reports generating thirty seconds of audio in half a second on a TPU-v4. That result identifies a tested setup and should not become a promise of identical latency on every processor, production assistant, or longer recording.