WavLate: first open-source multi-vector model for audio-text retrieval hits 37.9 Recall@1

antoine_chaffin · x · 2026-11-20

WavLate is claimed to be the first open late-interaction model for audio-text retrieval, keeping one vector per 0.31s of audio and per word. Without per-dataset fine-tuning it reaches a mean Recall@1 of 37.9 across AudioCaps and Clotho — best among nine open models evaluated. Weights, training data with mined hard negatives, and test pools are all released.

Original post →

More from Research

Research channel →