Silence Detection and Tail Feature Reset for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems using the Connectionist Temporal Classification (CTC) model require waiting for long periods of silence to display the last word of a previous section, leading to delayed text output when users are silent for an extended time or there are long silences in sentences.
Innovation Solution
Implementing a method that performs silence detection on the speech to be decoded, cuts off silent sections, resets tail features with preset silent frame features, and performs CTC decoding on the target speech to facilitate immediate display of text content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the CTC model waits for the CTC peak of the next word before displaying the current word, then the speech recognition accuracy is improved, but the text output delay increases when users are silent for extended periods
Solution Approach 1:
The system performs preliminary silence detection and speech cutting before CTC decoding. By detecting silent sections in advance and cutting the speech signal accordingly, the system prepares the processed speech segments beforehand, allowing the CTC model to decode and display text without waiting for extended silence periods while maintaining recognition accuracy through proper feature resetting.
Solution Approach 2:
The speech signal is segmented into active speech sections and silent sections. By dividing the continuous speech stream into discrete segments and processing only the relevant speech portions, the system eliminates unnecessary waiting time during silent periods while preserving the完整性 of speech content for accurate recognition.
2Reliability
If the system waits for the user to release the microphone button or for long silence to pass before displaying text, then the speech recognition completeness is improved, but the user experience responsiveness deteriorates
Solution Approach 1:
The system performs preliminary processing including silence detection, speech cutting, and tail feature resetting before CTC decoding. This preliminary preparation allows the system to display text promptly once processing is complete, improving responsiveness without sacrificing recognition completeness.
Solution Approach 2:
The system replaces the mechanical waiting approach (passively waiting for button release or silence) with an active signal processing approach. By using silence detection algorithms and automated speech cutting, the system actively identifies when speech ends and triggers text display, providing better user experience while maintaining complete recognition.
3Manufacturing precision
If the CTC decoding process waits for the next word's CTC peak before displaying the current word, then the word boundary accuracy is improved, but the display speed deteriorates during silent periods
Solution Approach 1:
The system performs preliminary speech cutting and tail feature resetting before CTC decoding. By preparing the speech segments in advance and resetting tail features to indicate speech termination, the system enables the CTC model to accurately identify word boundaries and trigger text display without unnecessary delays during silent periods.
Solution Approach 2:
The speech cutting module acts as an intermediary between the raw speech input and the CTC decoding process. It processes the speech signal by removing silent sections and resetting tail features, providing optimized input to the CTC model that improves both word boundary accuracy and display speed by eliminating ambiguous silent periods.
Data Source
AI summary
The present disclosure discloses a method including: performing a silence detection on a speech to be decoded; cutting the speech to be decoded off to obtain a target speech if detecting that the speech to be detected is a silent speech; resetting tail features of the target speech with preset tail features of silent frames; and performing a CTC decoding process on the target speech reset. In embodiments, when a large number of blank frames are carried in the speech to be decoded, the speech to be decoded is cut off, and the tail features of the target speech is placed with the tail features of the silent frames such that there may be one CTC peak when the CTC decoding process is performed on the tail features of the target speech. Therefore, a last word of text content may be displayed rapidly on a screen.