Technical report on StepAudio 3 Gen, a unified audio-generation system covering zero-shot TTS, voice design, vocal and sound-effect generation, and music, built on discrete autoregressive generation over residual vector quantization (RVQ) tokens instead of diffusion. Introduces a general audio tokenizer running at 12.5 Hz and a two-stage codebook-prediction architecture (time-axis autoregression for the first codebook, a lightweight causal Transformer for the rest). Reports state-of-the-art results on TTS and voice-design benchmarks, attributed to interference-aware progressive pretraining and RVQ-adaptor integration.
