Speech Recognition Model Training with Condition Number Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition models, particularly those using CNN-LTSM-CTC architecture, face processing latency issues due to high data read-write operations and parameter scale, making them unsuitable for embedded devices.
Innovation Solution
A lightweight method for training a speech recognition model is proposed, which includes a cascade-coupled architecture of convolution neural network, fully connected networks, and recurrent neural network. The model is trained with a constraint condition that limits the condition number of adjacent network layers, optimizing parameter matrices and reducing unnecessary normalization operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNN-LTSM-CTC model with residual network and normalization is used, then recognition accuracy is improved, but processing latency increases and it becomes unsuitable for embedded devices
Solution Approach 1:
The patent applies parameter changes by constraining the condition number of parameter matrices in the neural network layers. Specifically, it limits the condition number to be less than a predefined threshold (e.g., 10) through iterative optimization during training. This parameter constraint reduces the dynamic range of parameter values, eliminating the need for normalization operations and reducing processing latency while maintaining recognition accuracy on embedded devices.
2Measurement precision
If big parameter scale model with attention mechanism is used, then recognition accuracy is improved, but device complexity increases and it is only suitable for cloud end
Solution Approach 1:
The patent uses parameter changes by imposing condition number constraints on parameter matrices, which effectively reduces the actual number of independent parameters that need to be optimized. By constraining the condition number to be less than a threshold, the model achieves better generalization with fewer effective parameters, reducing model complexity while maintaining accuracy.
Solution Approach 2:
The patent segments the neural network into specific functional modules: convolutional feature extraction layers, recurrent sequence processing layers, and fully connected classification layers. This segmentation allows each module to be optimized independently with condition number constraints, reducing overall model complexity while maintaining the benefits of the attention mechanism for cloud-based deployment.
3Reliability
If normalization operations are performed multiple times, then model convergence is improved, but data read-write operations increase and processing efficiency decreases
Solution Approach 1:
The patent applies preliminary action by pre-constraining the condition numbers of parameter matrices during the training phase. By ensuring that parameter matrices satisfy the condition number constraint before deployment, the model achieves stable convergence without requiring multiple normalization operations during inference, thereby improving processing efficiency.
Solution Approach 2:
The patent changes the parameter distribution by applying condition number constraints during training, which transforms the parameter space to have more uniform scaling. This parameter transformation eliminates the need for repeated normalization operations, as the constrained parameters inherently have stable numerical properties that facilitate convergence without additional normalization steps.
Data Source
AI summary
This application relates to a method for training a speech recognition model comprising: providing a speech training data set comprising a plurality of speech data items and corresponding speech tags; providing a speech recognition model to be trained comprising a convolution neural network, a first fully connected network, a recurrent neural network and a second fully connected network which are cascade coupled together, wherein each of the networks comprises one or more network layers each having a parameter matrix; and the speech recognition model processing speech data items to generate corresponding speech recognition results; and using the speech training data set to train the speech recognition model such that the parameter matrices of at least two adjacent network layers satisfies a predetermined constraint condition; and the speech recognition model trained using at least one loss function can generate speech recognition results at an accuracy satisfying a predetermined recognition target.