Author | Lightnews

Ivan MH

@meresmanhiggs.bsky.social

36 followers 75 following 3 posts

Uruguayan engineer from Wales and Argentina, now in London. Signal processing and machine learning applied to music. RA at @c4dm. Ex @Ircam @IIE_FIng_UdelaR.

Posts Media Videos Starter Packs

Ivan MH @meresmanhiggs.bsky.social · Jun 25

Grateful to collaborators Mark Sandler and @emmanouilb.bsky.social, and to all the brilliant people at @c4dm.bsky.social for making these past months fun and so inspiring! Special thanks go to Adi (Aditya Bhattacharjee)🔥, with whom it's been (and continues to be) an absolute pleasure to work with🌟

Ivan MH @meresmanhiggs.bsky.social · Jun 25

Our lightweight framework - a GNN plus a cross-attention classifier - achieves state-of-the-art results with only 9% of the parameters of the previous SOTA 🎯❄️
We also extended the Sample100 dataset with precise timestamps & metadata!

Code & data: github.com/chymaera96/NeuralSampleID

GitHub - chymaera96/NeuralSampleID: An automatic sample identification (ASID) system using a contrastively trained GNN encoder.

An automatic sample identification (ASID) system using a contrastively trained GNN encoder. - chymaera96/NeuralSampleID

github.com

1 1

Ivan MH @meresmanhiggs.bsky.social · Jun 25

Our paper on automatic sample identification (arxiv.org/abs/2506.14684) was accepted at @ismir_conf 2025! 🎵🎶
We propose an architecture that can detect music samples that have been reused in new compositions, even after pitch-shifting, time-stretching, and other transformations!🧵

Image with the name of the paper "REFINING MUSIC SAMPLE IDENTIFICATION WITH A SELF-SUPERVISED GRAPH NEURAL NETWORK", names of the authors (Aditya Bhattacharjee, Ivan Meresman Higgs, Mark Sandler, and Emmanouil Benetos from Queen Mary University of London, UK), and a figure with the illustrated ASID methodology (2 stages): (A) Given a query, we compute segment-level embeddings (fingerprints), matched to reference embeddings via approximate nearest-neighbour (ANN) search; based on which, candidate songs are retrieved from the reference database through a lookup process (dotted arrows). (B) A multi-head cross-attention (MHCA) classifier refines and ranks candidates using node embedding matrices NMq (query) and NMr (references).

1 2 6

Reposted by Ivan MH

Robert Komaniecki @komaniecki.bsky.social · Nov 20

Reconnecting with the Twitter homies on Bluesky

240 4.7K 35K