Music Item Relation Determination Using Conditional Similarity Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to determine relations between music items in a database, especially as the database grows, since users cannot listen to all tracks to identify related items, and current solutions only apply to full audio files and not to both audio and symbolic tracks or submixes.
Innovation Solution
The use of conditional similarity networks (CSNs) to map music items into multiple vector subspaces based on different characteristics, allowing for the identification of relations across various types of music items, including audio and symbolic formats, by training models to minimize distance between similar items in each subspace.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the track database grows to include more music items, then the quantity of available tracks increases, but the ability to manually listen to and identify related tracks deteriorates
Solution Approach 1:
The patent replaces manual listening and identification processes with an automated machine learning system. The system uses neural networks to automatically analyze audio features, extract melodic representations, and determine relationships between tracks, eliminating the need for users to manually listen to each track to identify relations.
Solution Approach 2:
The system performs self-service by automatically analyzing and organizing tracks based on their audio characteristics. The machine learning models independently extract features, cluster similar tracks together, and provide recommendations without requiring user intervention or manual curation for each comparison.
2Measurement precision
If existing similarity methods are applied only to full audio files, then the processing accuracy is maintained, but the versatility of applying to individual tracks and submixes deteriorates
Solution Approach 1:
The patent segments the music analysis process into individual track-level processing units. Instead of treating only complete songs as the unit of analysis, the system extracts and analyzes individual tracks and submixes as separate entities, allowing the same similarity measurement methodology to be applied consistently across all levels of music organization.
Solution Approach 2:
The system achieves universality by designing a single framework that can handle multiple types of music items (individual tracks, submixes, and full songs) using the same neural network architecture and feature extraction methods. This multi-functional approach allows the system to maintain measurement precision across different scopes of music analysis.
3Reliability
If the system processes all tracks in the database, then comprehensive relations can be identified, but the processing time and computational resources increase
Solution Approach 1:
The system performs preliminary action by pre-processing and extracting audio features from all tracks during an initial phase. The neural network models are trained in advance on the entire database to recognize patterns and relationships. This preliminary processing enables the system to quickly identify relations between tracks during actual use without requiring real-time analysis of all possible combinations.
Solution Approach 2:
The system creates a simplified representation or 'copy' of each track in the form of extracted audio features and melodic representations. Instead of processing the complete audio files for every comparison, the system uses these condensed feature vectors to determine relationships, significantly reducing computational requirements while maintaining the essential information needed for relation identification.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of determining relations between music items, wherein a music item is a submix of a musical composition comprising one or more music tracks, the method comprising determining a first input representation for at least part of a first music item, mapping the first input representation onto to one or more subspaces derived from a vector space using a first model, wherein each subspace models a characteristic of the music items, determining a second input representation for at least part of a second music item, mapping the second input representation onto the one or more subspaces using a second model, and determining a distance between the mappings of the first and second input representations in each subspace, wherein the distance represents the degree of relation between the first and second input representations with respect to the characteristic modelled by the subspace.