Weighted Interpolation Language Model for Obscure Word Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition systems face challenges in accurately recognizing infrequent or obscure words due to low probability statistics models, leading to data offset issues and reduced recognition accuracy.
Innovation Solution
The method involves classifying raw speech corpus into categories, training language models for each category, implementing weighted interpolation to merge these models, and constructing decoding resources using both acoustic and interpolation language models to enhance recognition of obscure words by increasing their probability values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a universal language model is constructed using common speech corpus, then the model covers general speech patterns well, but the recognition accuracy for obscure words (medicine names, place names, etc.) deteriorates due to low probability statistics
Solution Approach 1:
The patent segments the speech corpus into multiple categories (news, weather, sports, finance, etc.) and constructs separate language models for each category. This segmentation allows the system to handle obscure words in specific domains by using the corresponding category-specific language model, thereby improving recognition accuracy for obscure words while maintaining manageable model construction complexity through modular organization.
Solution Approach 2:
The patent changes the parameter of language model probability distribution by introducing category-specific models with different probability distributions. Each category's language model has its own probability statistics tailored to that domain, allowing obscure words in specific categories to have higher probability values in their corresponding models, thus improving recognition accuracy without requiring a complete overhaul of the universal model.
2Measurement precision
If category-specific language models are constructed for different speech corpus categories, then the recognition accuracy for obscure words improves, but the device complexity increases due to multiple models and interpolation operations
Solution Approach 1:
The patent introduces an intermediary mechanism (weighted interpolation) that combines multiple category-specific language models. The interpolation weights dynamically select and blend probabilities from different category models based on the input speech context, providing a unified interface that hides the complexity of managing multiple models while achieving high recognition precision for obscure words through appropriate model selection and combination.
3Productivity
If the training corpus contains fewer obscure words, then the model training is simpler and faster, but the probability statistics for obscure words become insufficient leading to data offset and lower recognition accuracy
Solution Approach 1:
The patent performs preliminary action by pre-collecting and organizing domain-specific speech corpus for each category (news, weather, sports, finance, etc.) before model training. This preliminary corpus preparation ensures that each category-specific language model has sufficient obscure words relevant to its domain, enabling accurate probability statistics to be formed for obscure words without requiring excessive training data in the general corpus, thus maintaining training efficiency while improving reliability.
Data Source
AI summary
An automatic speech recognition method includes at a computer having one or more processors and memory for storing one or more programs to be executed by the processors, obtaining a plurality of speech corpus categories through classifying and calculating raw speech corpus; obtaining a plurality of classified language models that respectively correspond to the plurality of speech corpus categories through a language model training applied on each speech corpus category; obtaining an interpolation language model through implementing a weighted interpolation on each classified language model and merging the interpolated plurality of classified language models; constructing a decoding resource in accordance with an acoustic model and the interpolation language model; and decoding input speech using the decoding resource, and outputting a character string with a highest probability as a recognition result of the input speech.