Reproduce “Audio Fingerprinting with Holographic Reduced Representations” on OpenResearch

About this paper

Yusuke Fujita, Tatsuya Komatsu

This research introduces a method using Holographic Reduced Representations (HRR) to aggregate multiple neural audio fingerprints into a single composite representation, effectively reducing database storage requirements while maintaining high search accuracy and preserving the ability to pinpoint exact query locations. The approach achieved a Top-1 hit rate of 58.8% for 1-second queries with a 2x storage reduction factor, outperforming traditional aggregation techniques.

Key insights

• Holographic Reduced Representations enable the compositional encoding of sequential information into a single, compact vector, allowing multiple original fingerprints and their relative positions to be represented by one composite entry.
• The binding and summation properties of HRR allow for efficient search and retrieval of a specific component (query fingerprint at a specific relative position) from a composite fingerprint without needing to explicitly decode or recover the entire original sequence.
• While HRR significantly enhances aggregation performance, the limited gains from HRR-aware training suggest that the predominantly linear nature of HRR operations may present a bottleneck for further representational capacity improvements.

Summary by alphaXiv.