Speech Recognition Error Detection Data Generation via Mix-Up
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating learning data for speech recognition error detection face challenges in preventing overfitting due to the sensitivity of text data to token ID changes, which affects the continuity and meaning of sentences, making it difficult to effectively augment data for neural network training.
Innovation Solution
A method involving segmenting text data into token units, generating feature and label vectors, and performing mix-up operations, such as manifold-mix-up and shuffle permutations, to create additional training data that helps prevent overfitting and improve the robustness of neural network models for speech recognition error detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If text data is used for neural network training, then the model can learn speech recognition patterns, but changing token IDs slightly changes the text meaning completely, making data augmentation ineffective and causing overfitting
Solution Approach 1:
The patent applies parameter changes by transforming discrete token IDs into continuous feature vectors through embedding layers. This allows smooth interpolation and mixing of data points during augmentation, as the continuous vector space permits gradual transitions unlike discrete token IDs. The feature vectors are generated from text data and used for mix-up operations, enabling effective data augmentation while preventing overfitting.
Solution Approach 2:
The patent introduces feature vectors as an intermediary between text data and the neural network model. Instead of directly manipulating token IDs, the system converts text to feature vectors first, performs mix-up operations on these vectors, and then feeds the result to the model. This intermediary layer enables meaningful data augmentation by allowing continuous variations in the feature space while maintaining the original text semantics.
2Reliability
If mix-up operations are performed on feature vectors, then data diversity increases and overfitting is reduced, but the computational complexity of data generation increases
Solution Approach 1:
The patent applies preliminary action by pre-computing feature vectors from the training data before the actual mix-up operations. The embedding layer is trained in advance to convert text to feature vectors, and these pre-computed vectors are stored and reused during data generation. This preliminary processing reduces the computational burden during mix-up operations, as the system only needs to perform simple vector interpolation rather than complete data generation from scratch.
Data Source
AI summary
Disclosed is a method for generating data, the method is performed by one or more processors of a computing device. The method may include: segmenting text data generated based on speech information into a token unit; generating a first feature vector based on the text data segmented into the token unit, and generating a first label vector corresponding to the generated first feature vector, and generating a second feature vector and a second label vector by performing mix-up for each of the generated first feature vector and the generated first label vector.