LAION releases 1.48B-sample audio corpus; CLAP2-L hits 53.6% on MAEB-30
serrjoa · x · 2026-11-12
LAION released LAION-Audio-XL, a web-scale corpus of 1.48B collected audio samples curated into 365M audio-text pairs (1.39M hours, 707M captions), plus five CLAP2 models trained from scratch on a NaFlex-based encoder. Scaling studies reveal trade-offs between model size and data under fixed compute; CLAP2-L reaches 53.6% on MAEB-30, the best CLAP result on the leaderboard, outperforming same-size prior audio-text models. Dataset, captions, checkpoints, and code are all open-sourced.
More from Research
- PopVax Open-Sources a Personalized mRNA Cancer Vaccine Design for a Stage IV Patient — xennygrimmato_ · 2026-11-12
- Georgia Tech's Molei Tao Recruiting PhD Students and Postdocs in Diffusion Language Models — gautamcgoel · 2026-11-12
- HF researcher: report efficiency-gain metrics, not final benchmark scores — eliebakouch · 2026-11-12
- Researchers Push Efficiency-Gain Metrics Over Raw Benchmarks to Measure AI Progress — eliebakouch · 2026-11-12
- RoboCap-10k released: 10,000 hours of open egocentric robot data with metric actions — databoydg · 2026-11-12
- Google Research uses wearables for cardiovascular-kidney-metabolic risk profiling and coaching — Google Research · 2026-11-12