Word Category Estimation Using Graph Arc Scoring for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition techniques face challenges in accurately identifying named entities and maintaining high accuracy in the presence of speech recognition errors, particularly due to the limited features used and the complexity of language models.
Innovation Solution
A word category estimation apparatus and method that employs a probability model with multiple features, including an entire word category graph, to improve the robustness against speech recognition errors and enhance estimation accuracy by computing scores for arcs in the word category graph and re-performing speech recognition with vocabulary restrictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a language model with named entity class is used to identify named entities, then the extraction accuracy is improved, but the model complexity and number of parameters to be learned increase significantly
Solution Approach 1:
The patent segments the language model into two separate components: a standard language model for general word recognition and a named entity class model specifically for entity classification. This segmentation allows each model to be optimized independently, reducing the overall complexity while maintaining extraction accuracy.
Solution Approach 2:
The patent extracts the named entity class identification function from the general language model, creating a dedicated named entity class model. This extraction allows the system to use a simpler standard language model for general processing while applying the specialized entity model only where needed, reducing overall model complexity.
2Device complexity
If limited features (two preceding and succeeding words) are used in discriminative model, then the model simplicity is maintained, but the extraction accuracy decreases in sections with continuous speech recognition errors
Solution Approach 1:
The patent performs preliminary processing to generate a word graph from speech recognition results before applying the named entity class model. This preliminary action allows the system to identify and focus computational resources on segments containing speech recognition errors, improving accuracy without requiring complex features across the entire utterance.
Solution Approach 2:
The patent introduces a word graph as an intermediary structure between speech recognition and named entity extraction. This intermediary allows the system to propagate confidence information and contextual constraints through the graph, enabling accurate entity extraction even when local speech recognition features are limited or erroneous.
Data Source
AI summary
A word category estimation apparatus (100) includes a word category model (5) which is formed from a probability model having a plurality of kinds of information about a word category as features, and includes information about an entire word category graph as at least one of the features. A word category estimation unit (4) receives the word category graph of a speech recognition hypothesis to be processed, computes scores by referring to the word category model for respective arcs that form the word category graph, and outputs a word category sequence candidate based on the scores.