gwenlake

Lalala finds the song Shazam can't. Shazam needs the exact recording; Lalala works from what people actually have: a hummed tune, a few sung words, a vague memory ("the car ad with the whistling, around 2008"), a cover or live version, or a phone recording in a noisy bar. Instead of a lookup it runs an investigation: Llama 3.1 rewrites the memory into a sound query and a context query and keeps only grounded clues; several branches search in parallel (exact fingerprint, SAM Audio isolation as a second chance, PE-AV sound similarity, direct melody alignment of the hummed contour, one-way lyric hashes from Whisper, song cards written by Llama); a learned fusion votes, Llama reranks the top 10, and a calibrated confidence decides between one answer with a preview, a top-10 to listen to, or the question that narrows the search the most. Everything runs on one laptop (Apple Silicon, 24 GB) on Meta's open-weights models. Every claim is measured on frozen test sets the agent never sees: 93 % top-1 on noisy phone recordings, 19.6 % top-1 on 393 real hums (30× the published baseline), 30 % top-5 from a vague text memory. No audio and no lyrics are stored.