Audio Track Similarity Search Using FastMap and Symmetric KL Divergence

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current music recommendation algorithms, particularly those using Gaussian timbre representation and Kullback-Leibler divergence, face challenges in scalability due to high computational costs, non-metric properties, and the curse of dimensionality, making it difficult to handle large music collections effectively.

Innovation Solution

The method involves mapping audio tracks into a k-dimensional Euclidean vector space using the FastMap technique, selecting pivot points within median distances to preserve neighborhoods, and using the symmetric Kullback-Leibler divergence to ensure metric properties, allowing for efficient similarity determination and recommendation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If Kullback-Leibler divergence is used for computing acoustic music similarity, then recommendation precision is improved, but computational cost increases significantly

Engineering Contradiction:
Improverecommendation precisionVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The patent pre-computes and stores acoustic features (spectral centroid, spectral rolloff, spectral flux, zero crossings, beat histogram, pitch histogram) for all music tracks in the database before similarity queries are needed. This preliminary feature extraction allows the system to avoid expensive real-time KL divergence computations by first filtering candidates using simpler pre-computed representations, thereby reducing overall computational cost while maintaining precision.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary representation layer using simplified acoustic features and clustering approaches as a mediator between the raw audio data and the final KL divergence-based similarity ranking. This intermediary layer enables coarse filtering and preliminary sorting using less computationally intensive metrics, with KL divergence reserved for final verification on a reduced candidate set, thus balancing precision and computational efficiency.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If Kullback-Leibler divergence is used for similarity computation, then music recommendation accuracy is improved, but the system becomes difficult to scale to large music collections

Engineering Contradiction:
Improverecommendation accuracyVSAvoidscalability
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the music collection into clusters based on acoustic features (spectral centroid, spectral rolloff, spectral flux, zero crossings, beat histogram, pitch histogram) before performing similarity computations. This segmentation creates manageable groups of musically similar tracks, allowing the system to compute KL divergence only within or between small clusters rather than across the entire large database, thereby improving scalability while maintaining accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary clustering and feature extraction for all tracks in advance, organizing them into a structured representation that enables efficient querying. This preliminary organization allows the system to quickly identify candidate tracks for detailed KL divergence analysis without needing to compute similarities across all tracks simultaneously, thus enabling scaling to large collections while preserving recommendation accuracy.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If acoustic music similarity features with high degree of freedom are extracted, then similarity representation capability is improved, but indexing becomes difficult due to curse of dimensionality

Engineering Contradiction:
Improvesimilarity representation capabilityVSAvoidindexing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent extracts and selects only the most salient acoustic features (spectral centroid, spectral rolloff, spectral flux, zero crossings, beat histogram, pitch histogram) from the full audio signal, discarding redundant information. By focusing on these key features, the system reduces the dimensionality of the feature space while maintaining sufficient representation capability for accurate similarity computation, thereby simplifying indexing operations.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms the high-dimensional acoustic feature space into a lower-dimensional representation by computing summary statistics (mean, standard deviation) of the extracted features and using clustering algorithms that operate in this reduced space. This parameter transformation maintains the essential similarity information while significantly reducing dimensional complexity, making indexing and similarity search feasible for large-scale music collections.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP2273384B8A method and a system for identifying similar audio tracks
Publication Date: 2012.11.21 OSTE FORSCHUNGSINST FUR ARTIFICIAL INTELLIGENCE DER OSTEN STUDIENFUR KYBERNETIK

AI summary

In order to be able to handle very large numbers of tracks in a similarity determination in order to identify tracks similar to a predetermined track, a filtering method is used in order to identify a number of closest neighbour candidates between which the correct nearest neighbours are determined. Thus, the computationally heavy similarity determination is performed only on a subset of the tracks. This filtering step may be a fastmap determination of the tracks where the pivot points are determined not as the extreme points along the individual dimension but at the median thereof in order to avoid extremely high divergence values. This helps preserving the neighbourhoods. Also, the mapping is performed on the basis of a square-rooted Symmetric Kullback-Leibler (SKL) divergence which is more metric than the SKL and thus provides a better mapping.