Speech Recognition Delay Handling via Deep Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies fail to differentiate between temporary pauses and completion of utterances, leading to misrecognition and erroneous processing results, especially when delays occur during spoken input.
Innovation Solution
A speech processing method using a deep neural network model to recognize delays as incomplete utterances, temporarily store them, and combine with subsequent spoken inputs to form complete utterances, while deleting duplicate parts to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the speech recognition system processes utterances continuously without detecting delays, then processing speed is maintained, but recognition accuracy deteriorates due to misrecognizing delays as complete utterances
Solution Approach 1:
The system performs preliminary detection of delays during utterance reception and classifies them as incomplete utterances before final processing. This preliminary action prevents incorrect processing of incomplete utterances while maintaining efficient continuous processing capability.
Solution Approach 2:
The system dynamically adjusts its processing behavior based on real-time detection of delays versus complete utterances. When a delay is detected, the system pauses processing and waits for completion; when a complete utterance is detected, processing proceeds normally, optimizing both accuracy and efficiency.
2Reliability
If the system treats all pauses as complete utterances for simplicity, then processing is simplified, but harmful factors increase due to erroneous processing results
Solution Approach 1:
The system incorporates feedback mechanisms to detect delays in real-time during utterance reception. By continuously monitoring speech patterns and comparing them against learned models, the system identifies delays and adjusts processing accordingly, ensuring reliable results while maintaining manageable complexity through intelligent feedback loops.
3Loss of information
If duplicate utterance parts are not deleted, then processing is simpler, but information quality deteriorates due to redundant information in recognition results
Solution Approach 1:
The system extracts and removes duplicate utterance parts from the recognition results using a trained deep neural network model. This extraction process eliminates redundant information while preserving the essential meaning of the utterance, improving information quality without significantly increasing processing complexity.
Data Source
AI summary
Disclosed are a speech processing method and a speech processing apparatus in a 5G communication environment through speech processing by executing embedded artificial intelligence (AI) algorithms and/or machine learning algorithms. The speech processing method includes determining a temporary pause of reception of a first spoken utterance, outputting a first spoken response utterance as a result of speech recognition processing of a second spoken utterance received after the temporary pause, determining, as an extension of the first spoken utterance, a third spoken utterance that is received after outputting the first spoken response utterance, deleting, using a deep neural network model, a duplicate utterance part from a fourth spoken utterance that is obtained by combining the first and the third spoken utterance, and outputting a second spoken response utterance as a result of speech recognition processing of the fourth spoken utterance from which the duplicate utterance part has been deleted.