Speech Recognition Using Statistical Language Model Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech applications using the grammar method face challenges in recognizing natural speech due to the need for extensive interaction data collection, technical knowledge for statistical language model preparation, and costly rule descriptions for understanding utterance intentions, making it difficult for developers to recognize natural speech without significant effort and cost.
Innovation Solution
The solution involves using grammars as input data to generate a statistical language model for dictation recognition, converting recognition results into a form processable by the grammar method, and employing specific sentence generation and topic shifting to enhance search efficiency and reduce operational costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If speech recognition is performed by dictation using a statistical language model, then natural speech recognition capability is improved, but the need for collecting vast amounts of interaction data and preparing language models increases cost and complexity
Solution Approach 1:
The patent performs preliminary actions by collecting interaction data and preparing the statistical language model in advance, before the actual speech recognition service is deployed. This allows the system to handle natural speech recognition requests efficiently during operation without needing to perform complex data collection and model preparation at runtime, thus resolving the contradiction between natural speech capability and operational complexity.
2Measurement precision
If a statistical language model is used for natural speech recognition, then recognition accuracy for natural speech is improved, but the cost of data collection and model preparation increases
Solution Approach 1:
The patent implements self-service by automatically collecting interaction data from the speech application's operation logs and autonomously preparing the statistical language model using this collected data, without requiring manual intervention or external services. This automation significantly reduces the cost and effort associated with model preparation while maintaining high recognition accuracy.
3Measurement precision
If keywords are grouped to distinguish application-specific keywords from other words, then speech recognition processing accuracy is improved, but technical knowledge requirements and operational cost increase
Solution Approach 1:
The system automatically performs keyword grouping by analyzing the grammar structure and interaction data, identifying and categorizing application-specific keywords without requiring manual intervention. This automated approach maintains high processing accuracy while significantly reducing operational costs and the need for technical expertise.
4Adaptability or versatility
If rules are described to understand utterance intention from keyword sequences, then speech understanding capability is improved, but the complexity and cost of rule description increase
Solution Approach 1:
The patent replaces the mechanical approach of manually describing rules for understanding utterance intention with a statistical approach. By using the statistical language model to analyze patterns in interaction data, the system automatically learns to understand utterance intentions without requiring explicit rule descriptions, thus reducing complexity while maintaining or improving understanding capability.
Data Source
AI summary
A speech recognizing device. Natural speech recognizing means recognizes speech input in an application program by dictation. Recognition result converting means converts a recognition result from said natural speech recognizing means into a final recognition result processable by said application program on the basis of a grammar to he used for recognizing said input speech in a grammar method. The recognition result converting means further comprises candidate sentence generating means for evolving said grammar to generate candidate sentences that are candidates for said final recognition result: and matching means for selecting a candidate sentence as said final recognition result among the candidate sentences by matching said candidate sentences generated by said candidate sentence generating means against the recognition result by said natural speech recognizing means. The candidate sentence comprises a sentence having an associated highest score calculated according to the formula “(ngramWords−replacedWords−rejectedWords−insertedWords)/ngramWords.”