Deep Neural Network Speech Recognition Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems do not explicitly optimize model parameters for signal processing procedures, which limits their speech recognition performance.
Innovation Solution
Integrating a signal processing algorithm into a deep neural network (DNN) to convert and fuse signal processing and classification DNNs, allowing for explicit optimization of model parameters using back-propagation learning to minimize output errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If signal processing algorithm is used to extract feature parameters from speech input signal, then feature parameter extraction is achieved, but model parameters are not explicitly optimized for maximizing speech recognition performance
Solution Approach 1:
The patent merges the signal processing algorithm and classification algorithm into a unified deep neural network model. The signal processing DNN layers (including convolutional layers for feature extraction) are integrated with classification DNN layers, allowing both feature extraction and classification to be performed within a single end-to-end trainable system. This integration enables explicit optimization of model parameters for maximizing speech recognition performance while maintaining a structured approach to signal processing.
2Measurement precision
If signal processing algorithm is integrated into deep neural network, then model parameters can be explicitly optimized, but system complexity increases
Solution Approach 1:
The deep neural network is segmented into distinct functional modules: signal processing DNN layers (including convolutional layers for local feature extraction) and classification DNN layers. Each segment has a specific function - the signal processing segment extracts features from raw speech signals, while the classification segment performs recognition. This segmentation allows for explicit optimization of parameters in each segment while maintaining overall system manageability and interpretability.
Solution Approach 2:
The deep neural network structure serves multiple functions simultaneously: it performs signal processing (feature extraction), parameter optimization through back-propagation learning, and classification. The unified model handles both the extraction of feature parameters and the optimization of model parameters, reducing the need for separate processing stages and manual parameter tuning.
3Measurement precision
If back-propagation learning is used to learn model parameters, then speech recognition performance is maximized, but training time and computational resources increase
Solution Approach 1:
The signal processing DNN layers are initialized with parameters from a pre-trained model before fine-tuning. This preliminary action allows the system to start with reasonable parameter values and requires less training time to achieve optimal performance compared to training from random initialization. The pre-trained signal processing components provide a head start in feature extraction capabilities.
Solution Approach 2:
The training process uses dynamic learning rate adjustment and adaptive optimization techniques. The learning rate is adjusted during training based on performance metrics, allowing faster convergence in early stages and finer tuning in later stages. This dynamic approach optimizes the balance between training speed and convergence accuracy, reducing overall training time while maintaining high speech recognition accuracy.
Data Source
AI summary
Provided are a signal processing algorithm-integrated deep neural network (DNN)-based speech recognition apparatus and a learning method thereof. A model parameter learning method in a deep neural network (DNN)-based speech recognition apparatus implementable by a computer includes converting a signal processing algorithm for extracting a feature parameter from a speech input signal of a time domain into signal processing deep neural network (DNN), fusing the signal processing DNN and a classification DNN, and learning a model parameter in a deep learning model in which the signal processing DNN and the classification DNN are fused.