Research

Robust Neural Audio Fingerprinting using Music Foundation Models

Aug 2025 - Oct 2025

Fine-tuned music foundation models (MuQ, MERT) with a nonlinear projection head using contrastive learning. Achieved an 88.2% track-level hit rate and 95.5% segment-level F1, vs. 70.3% and 80.8% for the prior state of the art (BEATs), and beat the Shazam-like Dejavu by 39 points (49.2% hit rate). Co-first author, NeurIPS 2025 AI for Music Workshop.

SMucK: Symbolic Music in ChucK

Jan 2025 - Jun 2025

Designed a framework for symbolic music in the ChucK programming language. Accepted to NIME 2025.