Learned Filter Bank Layer for Speech Recognition Word Error Rate
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems rely on fixed mel-filter banks, which are not optimally designed for the task at hand, leading to suboptimal performance in word error rate (WER), and feature design is often separate from classifier design, resulting in features that may not be best suited for classification tasks.
Innovation Solution
Introducing a learned filter bank layer within a deep neural network that is optimized jointly with the network parameters using cross-entropy, allowing the filter bank to be adapted for the specific speech recognition objective, and incorporating delta learning and speaker adaptation techniques to improve feature extraction and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fixed mel-filter banks are used for feature extraction, then the system structure is simple and easy to implement, but the word error rate performance is suboptimal
Solution Approach 1:
The patent merges the feature extraction module (filter bank) with the classification module (neural network) into a unified end-to-end trainable system. The filter bank parameters are learned jointly with the neural network parameters through backpropagation, allowing the system to optimize both feature extraction and classification simultaneously for minimal word error rate.
Solution Approach 2:
The patent transforms the static fixed mel-filter bank into a dynamic learned filter bank where the parameters are adaptively adjusted during training. The filter bank parameters are updated via backpropagation based on the classification error, enabling the system to dynamically optimize feature extraction parameters for the specific speech recognition task.
2Reliability
If feature design is separated from classifier design, then the development process is simpler and more modular, but the features may not be optimally suited for the classification task
Solution Approach 1:
The patent combines feature design and classifier design into a single unified training process. The filter bank parameters and neural network parameters are optimized jointly through end-to-end backpropagation, ensuring that the extracted features are optimally suited for the classification task while maintaining manageable complexity through automated differentiation.
3Reliability
If learned filter banks are used instead of fixed mel-filter banks, then the word error rate improves by 5% relatively, but the training process becomes more complex and computationally intensive
Solution Approach 1:
The patent applies preliminary action by initializing the learned filter bank parameters with standard mel-filter bank values before training. This provides a good starting point for optimization, reducing the training time required to achieve convergence while still allowing the parameters to be refined for optimal word error rate performance.
Data Source
AI summary
Techniques for learning front-end speech recognition parameters as part of training a neural network classifier include obtaining an input speech signal, and applying front-end speech recognition parameters to extract features from the input speech signal. The extracted features may be fed through a neural network to obtain an output classification for the input speech signal, and an error measure may be computed for the output classification through comparison of the output classification with a known target classification. Back propagation may be applied to adjust one or more of the front-end parameters as one or more layers of the neural network, based on the error measure.