Semi-supervised NMF Acoustic Event Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing method for detecting acoustic events using Nonnegative Matrix Factorization (NMF) suffers from low identification accuracy, particularly in unknown environments, due to the presence of local solutions and the inability to accurately estimate basis matrices, leading to false identifications and incomplete sound source separation.
Innovation Solution
A signal processing device and method that incorporates a basis storage for acoustic event spectral bases and an identification model, using semi-supervised NMF to learn unknown spectral bases and calculate activation levels, thereby improving the accuracy of acoustic event identification by incorporating unknown acoustic event bases into the factorization process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If NMF is used for acoustic event detection and sound source separation, then the method can process acoustic signals and separate sound sources, but the identification accuracy is insufficient and false identifications occur in unknown environments
Solution Approach 1:
The basis matrix is segmented into two distinct parts: a teacher basis obtained from learning data representing known acoustic events, and an unknown basis that captures spectral patterns of unidentified sound sources. This segmentation allows the system to separately model known and unknown events, preventing false identifications while maintaining accurate detection of target events.
Solution Approach 2:
The unknown basis acts as an intermediary component that absorbs spectral patterns from unidentified sound sources in the mixture. By providing this intermediate representation, the system can accurately attribute spectral components to either known events (via teacher basis) or unknown sources (via unknown basis), thereby improving identification accuracy and reducing false positives.
2Manufacturing precision
If NMF is used to generate basis matrix from learning data, then sound source separation can be performed, but many local solutions exist and accurate basis matrix estimation fails
Solution Approach 1:
The basis matrix is divided into a teacher basis (from learning data) and an unknown basis (to be learned). This segmentation constrains the solution space by fixing the teacher basis, thereby avoiding the local solution problem that arises when the entire basis matrix is learned without constraints. The unknown basis is then learned only for spectral patterns not covered by the teacher basis.
Solution Approach 2:
The approach changes the parameter estimation strategy by separating fixed parameters (teacher basis from learning data) from learnable parameters (unknown basis). This parameter change transforms the optimization problem from estimating all basis parameters freely to estimating only the unknown basis parameters, thereby avoiding local solutions and improving estimation accuracy.
Data Source
AI summary
A signal processing device includes: a basis storage that stores an acoustic event basis group; a model storage that stores an identification model, as a feature amount, a combination of activation levels of spectral; an identification signal analysis unit that, upon input of a spectrogram of an acoustic signal for identification, performs sound source separation on the spectrogram by using a spectral basis set that is obtained by appending spectral bases corresponding to an unknown acoustic event that is an acoustic event other than the acoustic event specified as a detection target to the acoustic event basis group and causing only unknown spectral bases within the spectral basis set to be learned, and thereby calculating activation levels of spectral bases of the acoustic events in the spectrogram of the acoustic signal for identification; and a signal identify unit that identifies an acoustic event included in the acoustic signal for identification.