Utterance Verification Using Log-Likelihood Score Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current utterance verification techniques in speech recognition systems require significant computational resources and memory for model storage, and are not efficiently adaptable to different environments, which hampers their effectiveness and reliability.
Innovation Solution
A method and apparatus for utterance verification that processes log-likelihood scores from speech recognition to generate a verification score, comparing it with a threshold to accept or reject recognized vocabularies, using a reference score accumulator, verification score generator, and decision device, without the need for extensive additional computation or memory, leveraging log-likelihood scores from feature vectors and model vocabularies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional utterance verification techniques (Neural Network Model, garbage normalization, N-best based confidence) are used, then verification capability is provided, but computational resources and memory space requirements increase significantly
Solution Approach 1:
The patent extracts only the necessary log-likelihood scores from the speech recognition process that are already being computed, rather than implementing a complete separate verification system. By taking out and reusing these existing computational results, the system achieves verification capability without the full computational overhead of traditional methods like Neural Network Models
Solution Approach 2:
The log-likelihood scores computed for speech recognition are made multi-functional by reusing them for utterance verification as well. This universal approach allows the same computational results to serve dual purposes: both recognizing speech and verifying utterances, thereby eliminating redundant computations required by traditional verification techniques
2Reliability
If traditional utterance verification techniques are used, then verification capability is provided, but memory space for storing models increases
Solution Approach 1:
The patent extracts and reuses verification information from the existing speech recognition model outputs (log-likelihood scores), rather than storing separate verification models. This extraction approach provides verification capability without the memory burden of maintaining additional model structures like Neural Networks or garbage models
Solution Approach 2:
Instead of creating and storing separate verification models, the system copies and reuses the log-likelihood score information already generated during speech recognition. This copying approach allows verification to be performed using existing data structures without duplicating model storage requirements
3Reliability
If traditional utterance verification techniques are used, then verification is performed, but adjustment according to different environments is required
Solution Approach 1:
The patent employs parameter changes by adjusting the threshold value for verification decisions based on different environments and application requirements. This simple parameter adjustment mechanism provides environmental adaptability without requiring complex model retraining or multiple verification models, allowing the system to adapt to various conditions while maintaining verification accuracy
Data Source
AI summary
A method and apparatus for utterance verification are provided for verifying a recognized vocabulary output from speech recognition. The apparatus for utterance verification includes a reference score accumulator, a verification score generator and a decision device. A log-likelihood score obtained from speech recognition is processed by taking a logarithm of the value of the probability of one of feature vectors of an input speech conditioned on one of states of each model vocabulary. A verification score is generated based on the processed result. The verification score is compared with a predetermined threshold value so as to reject or accept the recognized vocabulary.