Language Model Look-Ahead Probability Calculation for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high computational cost and memory requirements of calculating language model look-ahead probabilities, especially for higher-order n-gram models, hinder efficient automatic speech recognition due to the large number of contexts and probabilities that need to be calculated and stored.
Innovation Solution
The method calculates look-ahead probabilities by utilizing a lower-order language model and updating only affected nodes when a higher-order model is used, leveraging the sparseness of n-gram probabilities and employing a back-off parameter to estimate probabilities, and using a multi-cache structure to efficiently generate and store higher-order probabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If higher order language model (e.g., trigram) is adopted to improve language model accuracy, then the recognition precision is improved, but the calculation cost increases dramatically due to the large number of different contexts that need to be calculated and stored
Solution Approach 1:
The patent segments the calculation process into two stages: first calculating lower-order LMLA probabilities (e.g., bigram) for all nodes in the search space, then selectively calculating higher-order LMLA probabilities (e.g., trigram) only for nodes where the lower-order calculation indicates potential improvement. This segmentation reduces the overall calculation burden while maintaining accuracy where needed.
Solution Approach 2:
Instead of calculating all higher-order LMLA probabilities exhaustively, the patent applies partial action by using lower-order probabilities as a filter to identify only those nodes where higher-order calculation would be beneficial. This avoids the excessive calculation cost of computing all possible trigram contexts while still improving accuracy at critical decision points.
2Measurement precision
If higher order language model is adopted to improve decoding accuracy, then the linguistic score precision is improved, but the memory requirements increase due to storing probabilities for all possible contexts
Solution Approach 1:
The patent segments the probability storage into hierarchical levels: lower-order probabilities (bigram) are stored for all nodes, while higher-order probabilities (trigram) are stored only for a subset of nodes identified as needing improved precision. This segmentation significantly reduces memory requirements while maintaining the ability to provide high-precision scores where necessary.
Solution Approach 2:
The patent applies partial storage action by maintaining complete lower-order probability tables for all contexts while storing only selected higher-order probabilities for nodes where they provide meaningful improvement. This avoids the excessive memory burden of storing all possible trigram contexts while preserving accuracy benefits at critical nodes.
3Measurement precision
If traditional method is used to calculate LMLA probabilities to ensure complete language model application, then the decoding accuracy is improved, but the processing time increases significantly making it unsustainable for real-time recognition
Solution Approach 1:
The patent performs preliminary calculation of lower-order LMLA probabilities for all nodes before attempting higher-order calculations. This preliminary action serves as a filter to identify which nodes actually require the computationally expensive higher-order calculations, thereby reducing overall processing time while maintaining decoding accuracy where it matters most.
Solution Approach 2:
Instead of applying the time-consuming higher-order calculation to all nodes, the patent applies it partially only to nodes where the lower-order probabilities indicate potential for improvement. This selective application maintains decoding accuracy for critical decisions while avoiding the excessive processing time that would result from universal higher-order calculation.
Data Source
AI summary
A system for calculating the look ahead probabilities at the nodes in a language model look ahead tree, wherein the words of the vocabulary of the language are located at the leaves of the tree,said apparatus comprising:means to assign a language model probability to each of the words of the vocabulary using a first low order language model;means to calculate the language look ahead probabilities for all nodes in said tree using said first language model;means to determine if the language model probability of one or more words of said vocabulary can be calculated using a higher order language model and updating said words with the higher order language model; andmeans to update the look ahead probability at only the nodes which are affected by the words where the language model has been updated.