ASR Decoding via Dynamic Beam Width Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large vocabulary automatic speech recognition (ASR) systems face challenges in real-time operation on battery-operated devices due to high computational loads, leading to increased power consumption and reduced battery life, as they require significant resources to handle computational peaks during utterance processing.
Innovation Solution
The implementation of inner frame histogram pruning, which dynamically updates the beam width during decoding by eliminating tokens that fall outside the updated beam width, reducing the variability of compute needed and lowering resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large vocabulary ASR systems are used to achieve high word recognition accuracy, then recognition quality is improved, but power consumption increases and battery life is reduced
Solution Approach 1:
The beam width is dynamically adjusted during decoding based on the histogram of token scores. The system transitions from a fixed beam width approach to a dynamic one where the beam width adapts to the distribution of token scores, allowing the system to maintain high recognition accuracy while reducing computational load and power consumption when full beam width is not necessary.
Solution Approach 2:
The system changes the beam width parameter during decoding based on the histogram analysis of token scores. By modifying this key parameter dynamically rather than keeping it fixed, the system can reduce computational requirements and power consumption while maintaining the necessary recognition accuracy for large vocabulary ASR.
2Measurement precision
If large vocabulary ASR systems are used to achieve high word recognition accuracy, then recognition quality is improved, but computational load increases making real-time operation difficult
Solution Approach 1:
The beam width is dynamically adjusted during decoding based on the histogram of token scores. The system transitions from a fixed beam width approach to a dynamic one where the beam width adapts to the distribution of token scores, allowing the system to maintain high recognition accuracy while reducing computational load and power consumption when full beam width is not necessary.
Solution Approach 2:
The system applies partial action by using histogram pruning to eliminate tokens that fall outside the updated beam width. Instead of processing all tokens with full beam width, the system selectively processes only the necessary portion of tokens, reducing computational load while maintaining sufficient recognition accuracy for real-time operation.
3Device complexity
If conventional beam search with fixed beam width is used, then implementation is simple, but computational peaks occur during utterance processing increasing power consumption
Solution Approach 1:
The beam width is dynamically adjusted during decoding based on the histogram of token scores. The system transitions from a fixed beam width approach to a dynamic one where the beam width adapts to the distribution of token scores, allowing the system to maintain high recognition accuracy while reducing computational load and power consumption when full beam width is not necessary.
Solution Approach 2:
The system performs preliminary action by updating the beam width histogram before processing each token. This preliminary update allows the system to identify and eliminate tokens that would fall outside the updated beam width, preventing unnecessary computational peaks and reducing power consumption before the actual token processing occurs.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A system, article, and method of automatic speech recognition with highly efficient decoding is accomplished by frequent beam width adjustment.