Speech Recognition Encoding via Streaming Fragment Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition devices with limited computing resources face high delays due to time-consuming encoding and decoding processes, which hinder efficient speech recognition.
Innovation Solution
Implementing a time-domain filtering scheme that allows for simultaneous acquisition and encoding of speech data, utilizing a streaming structure to reduce computational resources and delay by downsampling feature matrices, enabling real-time processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a whole sentence is acquired and input into the speech recognition model with non-streaming encoder and decoder, then the speech recognition accuracy is improved, but the encoding and decoding time increases significantly
Solution Approach 1:
The patent divides the speech data into speech fragments (e.g., 300 frames) that can be processed independently. The encoder processes each fragment separately to generate corresponding encoding results, enabling parallel processing and reducing overall encoding time while maintaining recognition accuracy through proper fragment segmentation.
Solution Approach 2:
The patent performs encoding operations in advance during the speech fragment acquisition phase. By encoding speech fragments as they are acquired rather than waiting for complete sentence input, the system prepares encoding results beforehand, reducing the total processing time when text conversion is needed.
2Reliability
If the encoder and decoder use non-streaming structure to ensure accurate processing, then the speech recognition reliability is improved, but the delay in speech recognition increases
Solution Approach 1:
The patent introduces a dynamic processing architecture where the encoder can operate in streaming mode on speech fragments while maintaining the reliability of non-streaming decoding. The system dynamically adjusts between fragment-based encoding and complete-sentence decoding, optimizing both speed and reliability based on processing needs.
Solution Approach 2:
By segmenting speech into manageable fragments and processing them through the encoder-streaming-decoder pipeline, the system achieves faster processing without compromising the reliability of the final text output, as each fragment is properly encoded and decoded through the reliable non-streaming decoder.
3Measurement precision
If the encoding process is performed after complete speech acquisition to ensure accuracy, then the recognition precision is improved, but the productivity of speech recognition decreases
Solution Approach 1:
The system performs preliminary encoding of speech fragments as they are acquired, preparing encoding results in advance. This preliminary action allows the decoder to work more efficiently when text conversion is needed, improving overall productivity without sacrificing recognition precision.
Solution Approach 2:
The patent enables continuous encoding of speech fragments during the acquisition phase, rather than waiting for complete speech input. This continuous useful action maintains productivity by keeping the encoder actively processing data throughout the speech acquisition period, while still ensuring accurate recognition through proper fragment handling.
Data Source
AI summary
A speech recognition method and apparatus, an electronic device, and a storage medium are provided. The speech recognition method includes: acquiring a to-be-recognized speech fragment; encoding acquired speech data while the to-be-recognized speech fragment is simultaneously acquired, to obtain an encoding result of the to-be-recognized speech fragment; and decoding the encoding result of the to-be-recognized speech fragment to obtain text data corresponding to the to-be-recognized speech fragment. This solution can reduce a delay of speech recognition in a speech recognition device having limited computing resources.