LSTM Acoustic Event Classification Using MOOD Cost Function
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional information classification systems, such as those used in speech recognition, are often complex and inefficient in terms of memory and processing power, making them impractical for low-power devices like mobile phones, and lack flexibility.
Innovation Solution
The use of a Long Short-Term Memory (LSTM) neural network with an improved cost function, specifically the many-or-one detection (MOOD) cost function, to efficiently train models for acoustic event classification, allowing the network to spike only when a confidence level is met during event detection, thus optimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional information classification systems are used for acoustic event detection, then classification capability is provided, but device complexity and resource consumption increase making them impractical for low-power devices
Solution Approach 1:
The audio signal is divided into discrete frames, with each frame independently processed by the neural network to determine presence/absence of the target acoustic event. This segmentation allows the complex classification task to be broken into manageable, resource-efficient units that can be processed sequentially on low-power devices.
Solution Approach 2:
The patent extracts only the essential features needed for binary classification (presence or absence of event) rather than performing full spectral analysis. The neural network is trained to extract only the critical patterns needed for detection, reducing computational complexity while maintaining accuracy.
2Reliability
If conventional information classification systems are used for acoustic event detection, then classification capability is provided, but memory and processing power requirements increase
Solution Approach 1:
The patent uses a lightweight neural network model that can be deployed on resource-constrained devices. The model is trained offline using comprehensive data, then deployed as a compact binary classifier that requires minimal memory and processing power during runtime, effectively using a 'cheap' computational approach for the detection task.
Solution Approach 2:
The patent transforms the continuous acoustic signal into discrete frame-based binary classification decisions. By changing the parameter representation from continuous spectral analysis to discrete presence/absence classification, the system reduces memory requirements and processing power while maintaining detection accuracy.
3Reliability
If the neural network is trained to spike all the times during event duration, then detection coverage is improved, but processing efficiency and resource usage deteriorate
Solution Approach 1:
The patent trains the neural network to detect events with partial action - it needs to spike at least once during the event duration rather than continuously throughout. This partial detection approach maintains reliability (the event is detected) while significantly improving processing efficiency by reducing the frequency of detection spikes and associated processing overhead.
Data Source
AI summary
An acoustic event detection and classification system includes a start-end point detector and multi-class acoustic event classification. A classification training system comprises a neural network configured to perform classification of input data, a training dataset including pre-segmented, labeled training samples, and a classification training module configured to train the neural network using the training dataset. The classification training module includes a forward pass processing module, and a backward pass processing module. The backward pass processing module is configured to determine whether a current frame is in a region of target (ROT), determine ROT information such as beginning and length of the ROT and update weights and biases using a cross-entropy cost function and a many-or-one detection (MOOD) cost function. The backward pass module further computes a soft target value using ROT information and computes a signal output error using the soft target value and network output value.