Speaker-Adaptive Speech Recognition Through Modular Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based speech recognition models often produce inaccurate results due to the lack of specificity in training data, failing to adapt to individual users' unique speech patterns and accents.
Innovation Solution
A speech recognition method and apparatus that includes a target feature extraction model bridged between a speech extraction model and a speech recognition model, allowing for speaker adaptive training by updating model parameters using user-specific speech data to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general AI-based speech recognition model is used, then the device can perform speech recognition without user-specific adaptation, but the recognition accuracy for individual users is poor
Solution Approach 1:
The speech recognition model is segmented into three distinct modules: speech extraction model, feature extraction model, and speech recognition model. This segmentation allows targeted updates to the feature extraction model using user-specific speech data without requiring retraining of the entire system, thereby improving recognition accuracy while controlling complexity.
Solution Approach 2:
The patent applies local quality by making the feature extraction model user-specific while keeping the speech extraction and recognition models general. This allows the system to adapt to individual user characteristics (local optimization) without requiring complete customization of the entire speech recognition pipeline.
2Measurement precision
If speaker adaptive training is implemented to improve recognition accuracy for individual users, then speech recognition accuracy improves, but the training complexity and computational resources required increase
Solution Approach 1:
The patent extracts the speaker-specific adaptation functionality into a separate feature extraction model that can be independently trained and updated. This extraction allows speaker adaptive training to be performed on a smaller, more manageable subset of the overall model, reducing training complexity while maintaining accuracy improvements.
Solution Approach 2:
The system performs preliminary speech extraction to obtain clean speech data from raw audio inputs before feeding it to the feature extraction model. This preliminary processing step prepares the data in advance, making the subsequent adaptive training more efficient and easier to implement.
3Measurement precision
If the feature extraction model is updated with user-specific speech data, then the model adapts to individual speech patterns improving accuracy, but the processing time and computational resources increase
Solution Approach 1:
The patent applies partial action by updating only the feature extraction model with user-specific speech data while keeping the speech extraction and recognition models fixed. This selective update approach reduces the computational burden and time required compared to training the entire model, while still achieving improved user-specific recognition accuracy.
Data Source
Figure 1~2
Figure 3~5
Figure 6
AI summary
A speech recognition method and apparatus, and a computer device and a storage medium, which relate to artificial intelligence. The speech recognition method comprises: acquiring first sample speech data corresponding to a target user and a first reference speech recognition result corresponding to the first sample speech data (S202); acquiring a target model before update, wherein the target model before update comprises a target speech extraction model, a target feature extraction model and a target speech recognition model, and the target feature extraction model is bridged between the target speech extraction model and the target speech recognition model (S204); inputting the first sample speech data into the target model before update to use the target speech extraction model, the target feature extraction model and the target speech recognition model to carry out speech recognition so as to obtain a first model output result (S206); obtaining, according to the first model output result and the first reference speech recognition result, a target model loss value corresponding to the target feature extraction model (S208); and updating, according to the target model loss value, a model parameter of the target feature extraction model in the target model before update to obtain an updated target model, so as to carry out speech recognition on speech data of the target user by using the updated target model (S210).