Speech Recognition Model Parameter Update via Word Error Rate Minimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing speech recognition models based on neural networks face limitations in precision improvement due to the use of different indicators and processing operations for learning and recognition.
Innovation Solution
A learning device that includes an extracting unit for feature extraction from speech data, a probability calculating unit for posterior probability calculation using a neural network-based speech recognition model, an error calculating unit for calculating word error rates, and an updating unit for parameter optimization to minimize loss, enabling precise learning of speech recognition models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If different indicators and processing operations are used for learning and recognition in neural network-based speech recognition models, then the model can be trained using standard methods, but the recognition precision cannot be improved further
Solution Approach 1:
The patent applies homogeneity by making the learning process and recognition process use the same indicators (word error rate) and processing operations (prefix searching). This unifies the two previously different processes, allowing the model to be trained with the same objective function used during recognition, thereby improving recognition precision without significantly increasing system complexity.
2Manufacturing precision
If prefix searching and word error rate calculation are used during learning, then recognition accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent implements self-service by having the learning device perform prefix searching and word error rate calculation using the same mechanisms as the recognition device. This allows the model to learn the actual recognition process dynamics, improving accuracy. The computational complexity is managed by reusing the same search and calculation frameworks for both learning and recognition.
Data Source
AI summary
A learning device (10) includes a feature extracting unit (11) that extracts features of speech from speech data for training, a probability calculating unit (12) that, on the basis of the features of speech, performs prefix searching using a speech recognition model of which a neural network is representative, and calculates a posterior probability of a recognition character string to obtain a plurality of hypothetical character strings, an error calculating unit (13) that calculates an error by word error rates of the plurality of hypothetical character strings and a correct character string for training, and obtains a parameter for the entire model that minimizes an expected value of summation of loss in the word error rates, and an updating unit (14) that updates a parameter of the model in accordance with the parameter obtained by the error calculating unit (13).