Aarush introduced WavLate on November 20, 2026, with a different approach to audio search: keep multiple numerical representations of a recording instead of squeezing the whole clip into one. The creator reports better average retrieval scores than eight other open models tested, and says model weights, training data and evaluation pools have been released. SourceSource
The job here is retrieval, not generating sounds or transcribing speech. According to Aarush’s Hugging Face community article, WavLate can rank recordings against a text description, or rank existing captions against a recording. Think searching a sound-effects library, not commissioning a new thunderstorm. Source
More room for the sounds
The launch article says WavLate retains 32 vectors per ten seconds of audio and one per caption token, a unit of text. Its scoring rule, MaxSim, selects each token’s best-matching audio-frame vector, then averages those matches. Different pieces of a description can therefore match different parts of a recording, rather than relying on one comparison between whole-input summaries. Source
The door slam and the background rain don’t have to share one numerical summary. The published configuration processes audio in ten-second windows, up to 30 seconds per clip. Source
An average win, not a clean sweep
In the creator’s tests across AudioCaps and Clotho, WavLate achieved 37.9% mean Recall@1 and 79.4% mean Recall@10, the highest averages among the nine open models evaluated. Recall@1 checks whether a correct match ranks first; Recall@10 checks whether one appears in the first ten results. These averages cover both search directions on both datasets, not general-purpose audio accuracy. Source
WavLate led every listed Clotho metric, but it didn’t evict the competition from every column. M2D-CLAP scored higher on AudioCaps audio-to-text Recall@1: 58.6%, against WavLate’s 53.6%. It also narrowly led AudioCaps text-to-audio Recall@10, at 88.4% versus 87.9%. Source
The evaluation used 964 AudioCaps test clips and 1,045 Clotho evaluation clips. The author says models were evaluated as released, without separate benchmark-specific fine-tuning, but all included AudioCaps and Clotho training splits in their training data. These were held-out test pools, not wholly unseen datasets. Source
The extra vectors need somewhere to live
In the author’s measurements, WavLate’s fp16 index occupied 7.9 KiB per clip, versus 2.0 KiB for MGA-CLAP. Scoring every caption against every clip in the AudioCaps test pool took 348 milliseconds, compared with 0.6 milliseconds for MGA-CLAP. Source
That timing used one A100 GPU in fp32 with already-decoded audio. It measures full-pool scoring, not the latency of one production query. Including audio encoding, text encoding and search, WavLate took 7.9 seconds, versus 6.5 seconds for MGA-CLAP. The enormous scoring gap is not the end-to-end gap. Source
Aarush’s launch article links the announced model and dataset releases. It reports no independent replication or timing on a production-scale audio collection. Source
Comments