Speech Recognition Handling Filled Pauses for Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems often misinterpret delays during utterance as completion, leading to incorrect processing results and inability to distinguish between temporary pauses and complete utterances, resulting in suboptimal performance.
Innovation Solution
A method and apparatus that determine temporary pauses within utterances by recognizing filled pauses, silent delays, or specific keywords, allowing for separate processing of incomplete and complete utterances, and maintaining an active speech recognition state to provide accurate results for both initial and additional utterances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If the speech recognition system processes utterances continuously without distinguishing pauses, then the response time is reduced, but the recognition accuracy deteriorates due to misinterpreting pauses as completions
Solution Approach 1:
The patent segments the utterance processing into distinct phases by detecting filled pauses (um, ah, er) and treating them as delimiters. This allows the system to separate incomplete utterances from complete ones, enabling accurate recognition while maintaining responsive processing through phased evaluation rather than continuous blind processing.
Solution Approach 2:
The system incorporates feedback by monitoring the speech stream for filled pause indicators and adjusting its processing state accordingly. When a filled pause is detected, the system pauses processing to wait for potential continuation, then resumes processing only when the utterance appears complete, thereby improving accuracy without significantly delaying response.
2Measurement precision
If the speech recognition system waits for complete utterances before processing, then the recognition accuracy is improved, but the user experience deteriorates due to delays and hesitations
Solution Approach 1:
The system performs preliminary processing by continuously analyzing the speech stream for filled pause indicators even before the utterance is fully delivered. This allows the system to be prepared to process the utterance immediately upon completion without requiring the user to wait for explicit completion signals, thereby improving user experience while maintaining accuracy through preliminary readiness.
Solution Approach 2:
The patent implements dynamic processing by adjusting the system's readiness state based on real-time detection of filled pauses. When a filled pause is detected, the system transitions to a waiting state; when the utterance continues without additional pauses, it transitions to a processing state. This dynamic adjustment optimizes both accuracy and user experience by matching processing timing to actual utterance completion.
3Productivity
If the speech recognition system processes additional utterances in the middle of reception, then the productivity is improved, but the processing complexity increases due to handling incomplete and complete utterances simultaneously
Solution Approach 1:
The patent segments the processing logic into distinct handlers: one for incomplete utterances (when filled pauses are detected) and one for complete utterances (when no filled pauses are present). This segmentation allows the system to process additional utterances efficiently without increasing overall complexity, as each segment follows a clear, defined processing path rather than requiring complex decision logic.
Solution Approach 2:
The system uses filled pause detection as an intermediary mechanism to determine processing behavior. Rather than directly handling the complexity of simultaneous incomplete and complete utterance processing, the filled pause detection acts as a simple mediator that directs the processing flow to appropriate handlers, thereby maintaining low complexity while enabling high productivity.
Data Source
AI summary
A speech recognition apparatus and an operating method thereof which execute a mounted artificial intelligence (AI) algorithm and/or machine learning algorithm to perform speech recognition and communicate with different electronic apparatuses and external servers in a 5G communication environment are disclosed. A speech recognition method according to an exemplary embodiment of the present disclosure includes determining a temporary pause for reception of a first utterance sentence in the middle of the reception of the first utterance sentence, outputting a speech recognition processing result of a second utterance sentence which is received after the temporary pause, separately from the first utterance sentence, determining a third utterance sentence which is received after outputting the speech recognition processing result of the second utterance sentence as an extension of the first utterance sentence, and outputting a speech recognition processing result of a fourth utterance sentence obtained by combining the first utterance sentence and the third utterance sentence. According to the present disclosure, a delay occurring in the middle of reception of uttering speech is recognized as an uncompleted utterance to be temporarily stored and a speech recognition processing result for an additional uttering speech received after the delay is provided and then uttering speech which is input again and the uttering speech before the delay are recognized as completed utterance and a speech recognition processing result is provided to improve the speech recognition processing performance.