Neural Network Auditory Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face inefficiencies in decorrelating spectral information from audio signals, leading to noise interference and suboptimal performance, particularly in handling rapid variations in auditory features.
Innovation Solution
The implementation of an artificial neural network that decorrelates spectral information using auditory filters, transforming it into cepstral information through connections between neuron layers, with weights determined by error-driven learning or offline optimization, and utilizing neuromorphic hardware for efficient processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional spectral analysis methods are used to extract auditory features, then the processing can be implemented with simple algorithms, but the signal-to-noise ratio is poor due to correlated noise in the power spectrum
Solution Approach 1:
The patent introduces an intermediary transformation process (cepstral transformation) that converts the correlated spectral domain into a decorrelated cepstral domain. This intermediary representation separates the useful auditory information from the correlated noise, achieving better signal-to-noise ratio while maintaining computational feasibility through established mathematical transformations.
Solution Approach 2:
The patent transforms the representation parameters from the spectral domain to the cepstral domain through mathematical transformation. This parameter change decorrelates the power spectrum components, converting correlated noise into independent cepstral coefficients, thereby improving the signal-to-noise ratio without requiring complex noise filtering algorithms.
2Reliability
If cepstral coefficients are used to decorrelate spectral information, then the signal-to-noise ratio is improved, but the computational complexity increases due to inverse Fourier techniques
Solution Approach 1:
The patent applies the cepstral transformation as a preliminary processing step before feature extraction and recognition. By pre-decorrelating the spectral information through inverse Fourier transformation, the system prepares the data in an optimal format for subsequent processing, improving overall system accuracy while maintaining efficient processing through the use of fast Fourier transform algorithms.
3Use of energy by moving object
If spiking neural networks are used for auditory feature extraction, then power consumption is reduced in neuromorphic hardware, but the implementation complexity of the frontend increases
Solution Approach 1:
The patent replaces traditional mechanical/computational signal processing systems with a spiking neural network implementation. This substitution leverages the event-driven, asynchronous nature of spiking networks to achieve lower power consumption in neuromorphic hardware, while the complexity is managed through biologically-inspired architectures that naturally handle temporal and spectral processing.
Data Source
AI summary
A system extracting features from a time-varying signal comprising a computer processor and a computer readable medium having computer executable instructions for providing: a bank of bandpass filters; a module approximating the output of those filters with nonlinear components; a module representing a decorrelated projection of the output of the filters with nonlinear components; and a module representing the temporal derivative of the decorrelated information with nonlinear components.