Neural Network Feature Extraction for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic speech recognition systems using Mel Frequency Cepstral Coefficients (MFCCs) struggle to differentiate between similar sounds, such as weak consonants, weak fricatives, and slurred speech, limiting their recognition accuracy.
Innovation Solution
The use of neural networks for non-linear feature extraction, optimizing a cost function from Linear Discriminant Analysis (LDA) to generate lower dimensional features that improve discriminative properties for speech recognition, by transforming MFCCs into neural network-LDA (NN-LDA) features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional MFCC features are used for speech recognition, then the system is simple to implement, but the recognition accuracy deteriorates for similar sounds
Solution Approach 1:
The patent transforms the feature extraction process by changing the mathematical parameters from linear MFCC operations to non-linear neural network operations. The neural network learns optimal non-linear transformations of MFCC features, enabling better differentiation of similar sounds while maintaining computational feasibility through gradient-based optimization.
Solution Approach 2:
The patent replaces traditional signal processing mechanics with neural network-based mechanics. Instead of using fixed linear transformations and hand-crafted features, the system employs learnable non-linear transformations that automatically adapt to the specific speech recognition task, improving accuracy for similar sound differentiation.
2Measurement precision
If non-linear feature extraction is used to improve sound differentiation, then recognition accuracy improves, but computational complexity increases
Solution Approach 1:
The patent applies partial non-linearity by using a shallow neural network with a limited number of layers and neurons. This provides just enough non-linear transformation to improve sound differentiation while avoiding the excessive computational cost of deeper networks. The model uses non-linear activation functions selectively to achieve the necessary discrimination power.
Solution Approach 2:
The patent optimizes computational efficiency by changing parameters such as network depth, width, and activation function choices. The system finds an optimal balance point where non-linear transformations provide sufficient sound differentiation without requiring excessive computational resources, making the approach practical for real-world deployment.
3Measurement precision
If linear discriminant analysis is used for feature transformation, then lower dimensional features are generated, but discriminative power is limited for similar sounds
Solution Approach 1:
The patent substitutes the linear algebraic operations of traditional LDA with neural network-based non-linear transformations. The neural network learns complex non-linear decision boundaries that better separate similar sounds in the feature space, replacing the limited linear projection of LDA with adaptive non-linear mappings that capture more nuanced acoustic differences.
Solution Approach 2:
The patent extends the LDA approach by introducing non-linear parameters through neural network activation functions. This transforms the fixed linear transformation of LDA into a learnable non-linear transformation that adapts to the specific characteristics of similar sounds, significantly improving discriminative power while maintaining dimensionality reduction benefits.
Data Source
AI summary
A system and method are presented for neural network based feature extraction for acoustic model development. A neural network may be used to extract acoustic features from raw MFCCs or the spectrum, which are then used for training acoustic models for speech recognition systems. Feature extraction may be performed by optimizing a cost function used in linear discriminant analysis. General non-linear functions generated by the neural network are used for feature extraction. The transformation may be performed using a cost function from linear discriminant analysis methods which perform linear operations on the MFCCs and generate lower dimensional features for speech recognition. The extracted acoustic features may then be used for training acoustic models for speech recognition systems.