GenAU
Snap's 1.25B-parameter transformer for ambient audio, trained on 57M captioned clips.
About
GenAU is Snap Research's take on scaling text-to-audio, and it attacks the data problem first. The June 2024 paper introduces AutoCap, an automatic audio captioning model, and uses it to build AutoReCap-XL: 57 million ambient audio clips with synthetic captions, about ninety times the scale of previous audio-text datasets. On top of that data it trains GenAU, a transformer-based generator scaled to 1.25 billion parameters.
Against state-of-the-art generators at similar size and data scale, the paper reports 4.7 percent better Frechet Audio Distance, 22.7 percent better Inception Score and 13.5 percent better CLAP score. Those are provider numbers, but the release is fully open: code, model checkpoints and the dataset are all public, which makes GenAU one of the more reproducible entries in audio generation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
FAD, IS and CLAP on AudioCaps vs SOTA baselines at similar scale (provider-reported).
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Snap Research paper (IJCV), public code, checkpoints and dataset, large-scale data contribution. Strengths: fully reproducible release, 57M-clip dataset, strong reported gains across three metrics. Docked for: (1) benchmarks are provider-reported; (2) ambient-sound focus rather than music. License: public code, checkpoints and dataset; check repo license before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Snap Research, arXiv 2406.19388, Jun 2024 (IJCV)
- · AutoCap captioning plus AutoReCap-XL: 57M captioned clips
- · 1.25B params; reported +4.7% FAD, +22.7% IS, +13.5% CLAP vs SOTA
- · Code, checkpoints and dataset public; Snap logo staged