Speech Recognition Re-ranking via Machine Learning Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition techniques in spoken chatting systems often generate erroneous responses due to misrecognized first-rank sentences, and there is a need for improved recognition results beyond simple processes.
Innovation Solution
A speech recognition method using machine learning that re-ranks multiple candidate sentences based on a learned model, selecting the highest-ranked candidate as the optimal result by generating learning data from feature extraction and correct answer sentences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition uses simple pattern matching methods, then the system is easy to operate and quick to respond, but the recognition accuracy deteriorates leading to erroneous responses
Solution Approach 1:
The patent applies preliminary action by pre-generating multiple candidate sentences with predicted ranks before final selection. The speech recognition system generates candidate sentences in advance with initial ranking, then uses a learning model to re-rank them. This allows the system to prepare multiple possibilities beforehand rather than relying on a single simple recognition pass, thereby improving accuracy without proportionally increasing operational complexity.
Solution Approach 2:
The patent introduces an intermediary element - a learning model (such as SVM or neural network) that acts as a mediator between the initial speech recognition results and the final selected sentence. This intermediary re-ranks the candidate sentences generated by the speech recognition unit, providing an additional layer of refinement that improves accuracy while keeping the overall system architecture manageable through modular design.
2Reliability
If speech recognition selects only the first-rank sentence, then the processing speed is fast and the system responds quickly, but the reliability deteriorates when the first-rank sentence is misrecognized
Solution Approach 1:
The patent applies partial action by not processing all candidate sentences equally, but rather focusing computational resources on re-ranking a limited set of top candidate sentences (e.g., top 3-5 candidates). This selective approach provides sufficient reliability improvement by examining multiple possibilities without the excessive time cost of analyzing all potential sentences, thus balancing reliability enhancement with acceptable processing time.
Solution Approach 2:
The system performs preliminary ranking and selection of candidate sentences before final decision-making. By pre-generating and pre-ranking multiple candidates, the system has already done part of the work needed for reliable selection, reducing the additional time required for verification and re-ranking while still improving upon single-rank selection reliability.
3Measurement precision
If the system generates multiple candidate sentences with ranking, then the recognition accuracy can be improved, but the device complexity increases due to additional processing units
Solution Approach 1:
The patent applies universality by designing the learning model to perform multiple functions: it re-ranks candidate sentences, learns from feedback about correct answers, and adapts to different speech patterns and contexts. This multi-functional component consolidates what could be multiple separate processing units into a single versatile system, improving ranking accuracy while limiting the increase in overall device complexity through functional integration.
Data Source
AI summary
Provided is a speech recognition method using machine learning, including: receiving a speech signal as an input, performing speech recognition to generate speech recognition result information including multiple candidate sentences and ranks of the respective candidate sentences; processing the multiple candidate sentences included in the speech recognition result information according to a machine learning model which is learned in advance and changing the ranks of the multiple candidate sentences to re-rank the multiple candidate sentences; and selecting the highest-rank candidate sentence among the re-ranked multiple candidate sentences as a speech recognition result. Particularly, the machine learning model is generated by: receiving the speech signal and a correct answer sentence as inputs; generating the speech recognition result information and a correct answer set; generating learning data by using the correct answer set; and performing the machine learning of changing the ranks of the candidate sentences.