Acoustic Model Adaptation via User Feedback Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated Search Recognition (ASR) engines face difficulties in accurately recognizing voice queries due to variations in language accents, leading to recognition errors.
Innovation Solution
Adapting acoustic models by selecting and refining voice queries based on user feedback, such as actions taken after transcription, to improve recognition accuracy and reduce errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If an initial acoustic model is used for voice query recognition, then the system can process queries, but recognition accuracy deteriorates due to language accent variations
Solution Approach 1:
The system performs preliminary adaptation of the acoustic model using bootstrap data and user feedback before actual deployment. Voice queries are pre-processed and used to refine the acoustic model parameters in advance, so that when the model encounters actual user queries with various accents, it is already optimized for better recognition accuracy
Solution Approach 2:
The system implements a feedback loop where user interactions with search results are captured and used to generate posterior recognition confidence measures. These feedback signals are continuously used to re-adapt and refine the acoustic model, improving its ability to handle accent variations over time while maintaining high recognition accuracy
2Reliability
If the acoustic model is adapted using all voice queries, then recognition accuracy improves, but computational resources and time are excessively consumed
Solution Approach 1:
Instead of using all voice queries for model adaptation, the system selectively uses a subset of queries that provide the most valuable training signals. By focusing on partial data that is most informative for accent variation, the system achieves good recognition accuracy improvement without the excessive computational time and resources that would be required to process all queries
Solution Approach 2:
The system performs preliminary filtering and selection of voice queries to identify those most suitable for adaptation. This pre-processing step identifies high-value training data before the actual model adaptation process, reducing the overall time required while maintaining effectiveness
3Reliability
If the acoustic model is adapted using all voice queries, then recognition accuracy improves, but system complexity increases
Solution Approach 1:
The system extracts and separates the model adaptation functionality from the main speech recognition system. By isolating the adaptation process as a distinct module that operates on selected subsets of data, the system improves recognition accuracy while avoiding the complexity increase that would result from integrating full-data adaptation into the core system
4Measurement precision
If user feedback is collected and processed, then posterior recognition confidence measures improve, but data processing requirements increase
Solution Approach 1:
The system extracts only the essential feedback signals needed for confidence measurement from the full user interaction data. By selectively processing only the relevant portions of user feedback rather than all generated data, the system achieves precise posterior recognition confidence measures while minimizing the quantity of data that needs to be processed
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for enhancing speech recognition accuracy. In one aspect, a method includes receiving voice queries, obtaining, for one or more of the voice queries, feedback information that references an action taken by a user that submitted the voice query after reviewing a result of the voice query, generating, for the one or more voice queries, a posterior recognition confidence measure that reflects a probability that the voice query was correctly recognized, wherein the posterior recognition confidence measure is generated based at least on the feedback information for the voice query, selecting a subset of the one or more voice queries based on the posterior recognition confidence measures, and adapting an acoustic model using the subset of the voice queries.