Speech Recognition Accuracy via Conditional Neural Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies face challenges in achieving high accuracy, especially in noisy environments and when the speech sequence contains syllables or words not covered by the language model, leading to decreased performance.
Innovation Solution
A method involving a speech recognition apparatus that uses both an acoustic model and a language model for initial text acquisition, with a secondary recognition process employing a neural network to improve accuracy by encoding and decoding the text based on feature vectors and hidden states, allowing for robust noise handling and expanded vocabulary recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If only acoustic model and language model are used for speech recognition, then the system complexity is low, but the recognition accuracy decreases in noisy environments and for unseen syllables
Solution Approach 1:
The system segments the speech recognition task into multiple stages: first using acoustic model and language model for initial text acquisition, then conditionally applying neural network-based sequence-to-sequence model for correction when accuracy is insufficient. This segmentation allows the system to achieve high accuracy only when needed, rather than always using the complex neural network approach.
Solution Approach 2:
The system dynamically adjusts its processing path based on the calculated accuracy of the first text. When accuracy meets the threshold, the simpler acoustic model approach is sufficient; when accuracy is below the threshold, the system transitions to the more complex neural network correction path. This dynamic adaptation optimizes the balance between accuracy and complexity.
2Measurement precision
If neural network-based sequence-to-sequence model is always used for correction, then the recognition accuracy improves, but the processing time increases
Solution Approach 1:
The system applies the computationally intensive neural network correction only partially - specifically, only when the initial recognition accuracy is below the threshold. This partial application of the correction mechanism avoids the excessive processing time that would result from always applying the neural network, while still achieving accuracy improvement when needed.
Solution Approach 2:
The system uses feedback from the accuracy calculation to control whether the correction process is executed. The accuracy of the first text serves as feedback that determines the next action: if accuracy meets the threshold, no correction is needed; if accuracy is below the threshold, correction is applied. This feedback mechanism prevents unnecessary processing time.
3Adaptability or versatility
If acoustic model and language model are used, then the system structure is simple, but the vocabulary coverage is limited to trained syllables
Solution Approach 1:
The neural network-based sequence-to-sequence model serves as a universal component that can handle both correction of recognition errors and recognition of unseen syllables/words. This multi-functional approach allows the system to expand its vocabulary coverage beyond what the acoustic model and language model can handle, without requiring separate specialized components for each function.
Data Source
AI summary
Disclosed is a speech recognition method and apparatus, the method including two recognition processes, a first recognition process being performed using an acoustic model and a language model and a second recognition process being performed without distinguishing between the acoustic model and the language model in response to an accuracy of a result of the first recognition process not meeting a threshold. The apparatus including a processor configured to acquire a first text from a speech sequence using an acoustic model and a language model, determine whether an accuracy of the first text meets a threshold, and acquire a second text from the first text based on a parameter generated in acquiring the first text, in response to the accuracy of the first text being below the threshold.