Expected-Response Confidence Tuning in Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face issues such as false rejections, false acceptance, and out of vocabulary insertion errors, particularly when different users with varying speech patterns interact with the system, leading to inaccuracies in interpreting speech.
Innovation Solution
Adapting speech recognition confidence scores using expected responses to adjust confidence score values, reducing the risk of false rejections and improving accuracy by comparing recognition hypotheses to expected responses and updating statistical data accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition systems use fixed confidence score thresholds, then the system structure remains simple, but accuracy deteriorates due to false rejections and false acceptances across different users
Solution Approach 1:
The system performs preliminary adaptation of confidence scores using expected responses before final recognition decisions are made. Statistical data from expected responses is used to pre-adjust confidence scores, allowing the system to account for user-specific speech patterns in advance, thereby improving accuracy without requiring complex real-time adjustments for each user
Solution Approach 2:
The system dynamically changes confidence score parameters based on statistical data derived from expected responses. By adjusting confidence scores as a parameter rather than using fixed thresholds, the system adapts to different users and speech patterns, resolving the contradiction between maintaining simple system structure and achieving high recognition accuracy
2Measurement precision
If the system adapts confidence scores for each user, then recognition accuracy improves, but the complexity of data processing and model adjustment increases
Solution Approach 1:
The system performs preliminary adaptation of confidence scores using expected responses before final recognition decisions are made. Statistical data from expected responses is used to pre-adjust confidence scores, allowing the system to account for user-specific speech patterns in advance, thereby improving accuracy without requiring complex real-time adjustments for each user
Solution Approach 2:
The system introduces statistical data derived from expected responses as an intermediary between raw speech input and recognition decisions. This intermediary layer processes user-specific patterns in a standardized way, simplifying the overall data processing complexity while still enabling user-specific adaptation
3Productivity
If the system uses standard speech recognition without adaptation, then processing speed remains fast, but false rejections increase requiring repeated utterances
Solution Approach 1:
The system performs preliminary adaptation of confidence scores using expected responses before final recognition decisions are made. This pre-adaptation reduces false rejections by accounting for user-specific patterns in advance, thereby improving recognition reliability without requiring repeated utterances that would reduce productivity
Solution Approach 2:
The system uses expected responses as feedback to adjust confidence scores. By incorporating this feedback mechanism, the system learns from correct responses and improves its reliability over time while maintaining fast processing speeds, thus resolving the contradiction between productivity and reliability
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of the disclosure provide for adapting speech recognition confidence scores to one or more users based on expect response(s). Some embodiments, receive input speech and generate, based on the input speech and using a recognition algorithm, a recognition hypothesis. The recognition hypothesis may comprise at least one pairing comprising a word and an initial confidence score associated with the word. Some embodiments, determine whether the recognition hypothesis matches an expected response based at least in part on comparing the word to a corresponding word in an expected response. Some embodiments, in response to determining that the recognition hypothesis matches the expected response, updates statistical data associated with the word.