Neural Network Speech Decoding With Stable Token Reuse
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems struggle with responsiveness and efficiency in real-time decoding, particularly when performing partial decoding of speech inputs, leading to inaccurate intermediate results and increased processing time due to repeated decoding steps.
Innovation Solution
A neural network-based decoding method that reuses tokens from previous decoding steps by identifying a stable time based on matching substrings and using these tokens for subsequent decoding, reducing the number of decoding steps through a relay and limited relay decoding approach.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If repeated decoding steps are performed for partial speech inputs, then decoding completeness is improved, but processing time increases and responsiveness deteriorates
Solution Approach 1:
The system performs preliminary decoding on partial speech inputs to generate intermediate results. These preliminary decoding steps create candidate beams that can be reused in subsequent decoding, avoiding the need to reprocess the entire speech sequence from scratch. This preliminary action on partial inputs enables faster overall processing while maintaining accuracy.
Solution Approach 2:
The decoding process is segmented into multiple partial decoding steps, where each step processes a portion of the speech input independently. By dividing the complete decoding task into segments that can be performed progressively, the system achieves real-time responsiveness while ensuring complete decoding accuracy through the accumulation of intermediate results.
2Manufacturing precision
If complete decoding is performed on the entire speech input, then decoding accuracy is improved, but responsiveness to partial speech inputs deteriorates
Solution Approach 1:
The system performs partial decoding actions on incomplete speech inputs rather than waiting for complete input. By executing decoding operations on partial sequences and using the resulting candidate beams for subsequent steps, the system achieves responsiveness to partial inputs while maintaining final accuracy through the cumulative effect of multiple partial decoding operations.
3Manufacturing precision
If multiple decoding steps are performed on partial speech inputs, then decoding completeness is improved, but processing load increases
Solution Approach 1:
The system recovers and reuses candidate beams from previous decoding steps instead of discarding them. By storing the candidate beams generated in partial decoding steps and reusing them in subsequent decoding operations, the system avoids redundant processing and reduces the overall computational load while maintaining decoding completeness.
Data Source
AI summary
A decoding method includes receiving an input sequence corresponding to an input speech at a current time; and in a neural network (NN) for speech recognition, generating an encoded vector sequence by encoding the input sequence, determining reuse tokens from candidate beams of two or more previous times by comparing the candidate beams of the previous times, and decoding one or more tokens subsequent to the reuse tokens based on the reuse tokens and the encoded vector sequence.