Music Analysis Apparatus Using Probabilistic Model for Robust Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing music search technologies are not robust against errors in designated note sequences, particularly when wrong notes or sequences differently arranged from reference music pieces are provided, leading to inaccurate searches.
Innovation Solution
A music analysis apparatus that generates a time series of feature values from designated notes and computes an evaluation index value using a probabilistic model learned from reference music pieces, incorporating duration features and pitch information to assess the probability of sequence similarity across multiple unit periods, allowing for robust music search even with errors in the designated sequence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exact sequence matching is used to search music pieces, then search precision is improved, but search reliability deteriorates when errors exist in designated notes
Solution Approach 1:
The patent transforms the exact sequence matching approach into a probabilistic matching approach by changing the parameter from binary match/no-match to probability-based similarity scoring. The system calculates the probability that each reference music piece contains the designated sequence of notes, allowing for robust search even when errors exist in the designated notes.
Solution Approach 2:
The patent introduces a probabilistic model as an intermediary between the designated sequence of notes and the reference music pieces. This intermediary layer computes similarity probabilities, enabling the system to handle errors in the designated notes while maintaining search accuracy.
2Reliability
If probabilistic model with machine learning is applied, then search robustness is improved, but device complexity increases
Solution Approach 1:
The patent performs preliminary machine learning training offline to generate the probabilistic model before actual music search operations. This preliminary action separates the complex model training process from the real-time search process, reducing the complexity burden on the search system itself while maintaining robust search capabilities.
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
A music analysis apparatus has a feature extractor and an analysis processor. The feature extractor generates a time series of feature values from a sequence of notes which is designated as an object of analysis. The analysis processor computes an evaluation index value which indicates a probability that the designated sequence of notes is present in each of a plurality of reference music pieces by applying a probabilistic model to the time series of the feature values generated from the designated sequence of notes. The probabilistic model is generated by machine learning of the plurality of reference music pieces using time series of feature values obtained from the reference music pieces.