Skip to content

SoundStorm

Parallel audio token generation for faster synthesis

Visit official site

What SoundStorm is

The technique is described as a Google Research audio generation approach that reduces synthesis time by producing audio tokens in parallel instead of step by step.

What you can do with it

  • Generate audio tokens in parallel
  • Keep quality aligned with the original audio signal
  • Control speech content and speaker traits

How it works

  • A neural audio codec compresses the waveform
  • A Transformer based sequence-to-sequence model is used
  • Tokens are predicted from coarse to fine levels

Pros & cons

✓ What we like

  • Long-sequence inference time is reduced
  • Long text speech and natural dialogue are supported

! What to watch out for

  • Generation starts from text or audio prompts
  • The described pipeline uses a neural audio codec

FAQ

What problem does it address?

The material states slow inference of autoregressive models on long sequences.

What can it control?

The material names speech content, speaker voice and speaker turns.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review