Audio Event Detection Using Mel-Scale Filter Banks and NMF
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio event detection systems require large amounts of training data and rely heavily on neural networks, which can be inefficient and computationally intensive, especially in real-time applications.
Innovation Solution
An audio processing system that uses a feature extraction block with a filter bank spaced according to the mel-frequency scale to derive spectral features, followed by a classification unit that performs non-negative matrix factorization (NMF) to compare input signals with a dictionary, allowing for efficient detection of audio events by representing signals as a weighted sum of dictionary features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural networks are used for audio event detection, then detection accuracy is improved, but computational complexity and training data requirements increase
Solution Approach 1:
The patent segments the audio signal processing into distinct stages: filter bank decomposition into frequency bins, followed by Non-Negative Matrix Factorization (NMF) to separate source signals. This segmentation replaces the monolithic neural network approach with modular, interpretable processing steps that reduce computational complexity while maintaining detection accuracy.
Solution Approach 2:
The patent introduces an intermediary representation through the use of spectrograms and NMF-derived source estimates as intermediate signals. These intermediaries bridge the raw audio input and final event detection, allowing for more efficient processing compared to direct neural network mapping, while preserving the information needed for accurate classification.
2Measurement precision
If neural networks are used for audio event detection, then detection accuracy is improved, but processing time increases
Solution Approach 1:
The patent employs periodic frame-based processing where the audio signal is divided into overlapping frames, and NMF is applied periodically to each frame. This periodic action enables real-time processing with controlled computational load, contrasting with the continuous heavy computation required by neural networks, thus reducing overall processing time while maintaining accuracy.
Solution Approach 2:
The patent changes the processing parameters by using fixed dictionary matrices in NMF that are pre-computed from training data, rather than performing full neural network inference at runtime. This parameter change allows for faster real-time processing since the computationally intensive parts are done offline, reducing online processing time while preserving detection accuracy.
3Reliability
If large amounts of training data are used, then model performance is improved, but data storage and processing requirements increase
Solution Approach 1:
The patent extracts only the essential statistical properties of the training data by computing fixed NMF dictionaries that capture the dominant source patterns. Instead of storing large amounts of raw training data, only the compact dictionary matrices are retained, dramatically reducing storage requirements while maintaining model performance through the preserved structural information.
Solution Approach 2:
The patent replaces the need for large persistent training datasets with compact, disposable dictionary representations. These dictionaries are computed once from training data and then used repeatedly for detection without requiring the original training data to be stored, effectively discarding the large training datasets after extraction of essential features, thus reducing long-term storage requirements.
Data Source
AI summary
An audio processing system is described for an audio event detection (AED) system. The system includes a feature extraction block configured to derive at least one feature which represents a spectral feature of the input signal.