Pruned Language Model Key Phrase Recognition via Incentive Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current language recognition models used in speech recognition technology, obtained by pruning large language models, often result in decreased accuracy due to the loss of important edge weights during the pruning process.
Innovation Solution
The method involves adjusting the probabilities of relationships between elements in a language recognition model based on a textual segment, by loading a key phrase language model and a large language model, extracting reference edges, updating weights, and configuring these weights as incentive weights in the language recognition model, to enhance the recognition of key phrases during speech recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a large language model is pruned to obtain a language recognition model, then the model size is reduced and processing speed is improved, but the accuracy of speech recognition decreases due to loss of important edge weights
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing incentive weights for key phrase edges in the pruned language model before speech recognition occurs. During recognition, these pre-computed weights are directly applied to enhance key phrase recognition without affecting the overall model structure or requiring re-pruning, thus maintaining both speed and accuracy improvements.
2Measurement precision
If the language recognition model is adjusted to enhance key phrase recognition, then the accuracy of key phrase recognition is improved, but the complexity of the model increases
Solution Approach 1:
The patent applies local quality by selectively enhancing only the edge weights corresponding to key phrase paths in the language model, while leaving other edges unchanged. This localized adjustment focuses computational resources on improving key phrase recognition accuracy without unnecessarily complicating the entire model structure, thereby improving accuracy with minimal increase in overall model complexity.
Data Source
AI summary
Embodiments of this disclosure disclose a speech recognition method, apparatus, and device, and a storage medium. The method in the embodiments of this disclosure includes: adjusting a probability of a relationship between at least one pair of elements in a language recognition model according to a probability of the relationship between the at least one pair of elements in a textual segment; inputting a to-be-recognized speech into a speech recognition model including the language recognition model; and determining, according to the adjusted probability of relationship between the at least tone pair of elements in the language recognition model, a sequence of elements corresponding to the to-be-recognized speech as a speech recognition result.