Speech Recognition Language Model Customization via Phrase Tokenization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies face challenges in accurately transcribing phrases due to variations in pronunciation, silent elements, and context-based pronunciations, leading to inefficient customization of language models and excessive permutations of possible pronunciations.
Innovation Solution
A system and method for customizing language models by generating special pronunciations for phrase components, tokenizing phrases, and using a pron guesser to determine pronunciations, which are then combined and added to the language model, while considering initial and non-initial tokens and silent elements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If phrase pronunciations are created by combining pronunciations of component elements, then the language model can be customized with site-specific names and phrases, but the system generates excessive permutations of possible pronunciations leading to inefficiency
Solution Approach 1:
The patent segments the phrase processing into distinct components: tokenizing the phrase into individual elements, identifying which tokens are pron components, and separately determining pronunciations for each. This segmentation allows the system to focus computational effort only on tokens that require pronunciation generation, rather than processing all tokens uniformly, thereby reducing excessive permutations while maintaining customization capability.
Solution Approach 2:
The patent applies different processing rules to different parts of the phrase based on their characteristics. Specifically, it identifies pron components (tokens whose pronunciations differ when occurring in a phrase) and applies specialized pronunciation determination logic only to those tokens, while other tokens use standard pronunciation lookup. This localized approach reduces overall computational complexity while maintaining accuracy for critical elements.
2Measurement precision
If the system considers context-based pronunciations for pron components, then phrase recognition accuracy improves, but the complexity of determining pronunciations increases
Solution Approach 1:
The patent performs preliminary identification of pron components during the tokenization phase, before pronunciation determination begins. By pre-marking which tokens require special pronunciation handling based on their identity as pron components, the system prepares the data structure in advance, allowing the pronunciation determination step to focus only on these identified elements rather than evaluating all tokens, thus reducing overall complexity while maintaining accuracy.
3Measurement precision
If silent elements and initial/non-initial tokens are considered in pronunciation generation, then transcription accuracy improves, but the processing time and computational effort increase
Solution Approach 1:
The patent applies specialized processing rules specifically to initial tokens and pron components, while using standard processing for other tokens. Initial tokens receive special attention for pronunciation determination, and silent elements are handled with optimized logic that avoids full pronunciation generation. This localized differential processing maintains high transcription accuracy for critical elements while reducing overall processing time by avoiding unnecessary computation on non-critical tokens.
Data Source
AI summary
A system and method for a speech recognition technology that allows language models to be customized through the addition of special pronunciations for components of phrases, which are added to the factory language models during customization. It allows components of a phrase to have different pronunciations inside customer-added phrases than are specified for those isolated components in the factory language models.