Cloud Speech Recognition With Relationship-Based Hot Word Lists
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Universal speech recognition models have low accuracy for special words like person names, product names, or service keywords, and using a large hot word list to improve recognition efficiency leads to reduced speech recognition speed.
Innovation Solution
Generate a hot word list by combining sublists based on different relationship features of a user, such as personal networks and group affiliations, to prioritize relevant words and reduce invalid entries, thereby improving recognition efficiency without compromising accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large hot word list is used to improve recognition accuracy for special words, then speech recognition accuracy is improved, but speech recognition speed decreases
Solution Approach 1:
The hot word list is segmented into multiple hot word sublists, where each sublist corresponds to a specific relationship feature (e.g., personal relationship network, workgroup, company). During speech recognition, only the sublist corresponding to the current relationship feature is activated, rather than processing the entire hot word list. This segmentation reduces the number of hot words that need to be matched in real-time, thereby improving speech recognition speed while maintaining accuracy for relevant special words.
2Measurement precision
If all hot words are included in the hot word list to ensure comprehensive coverage, then recognition accuracy for various special words is improved, but the complexity of the system increases
Solution Approach 1:
The hot word list is transformed from a static, comprehensive list into a dynamic structure where different sublists are activated based on the current relationship feature. The system dynamically selects and activates only the relevant hot word sublist corresponding to the current interaction context, rather than maintaining a fixed large list. This dynamic approach reduces system complexity by loading only necessary data into memory during each recognition task.
Solution Approach 2:
Hot word sublists are pre-organized and prepared according to different relationship features before the speech recognition task. Each sublist is预先 (pre) constructed and stored separately, so that during actual speech recognition, the system only needs to select and activate the appropriate pre-prepared sublist rather than processing a comprehensive list from scratch. This preliminary organization reduces real-time processing complexity.
Data Source
AI summary
A speech recognition method and apparatus are provided. The method is applied to a cloud management platform. The speech recognition method includes: The cloud management platform obtains a to-be-recognized speech of a user; the cloud management platform obtains a hot word list of the user, where the hot word list of the user is generated by combining a plurality of hot word sublists, and different hot word sublists correspond to different relationship features of the user; and the cloud management platform performs speech recognition on the to-be-recognized speech based on the hot word list. The hot word list is generated by combining the plurality of hot word sublists corresponding to the different relationship features of the user. Therefore, invalid hot words are reduced, and speech recognition efficiency can be improved without reducing recognition accuracy.