Posterior Feature Partial Distance Elimination Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high computational costs associated with evaluating posteriors in automatic speech recognition systems, particularly due to the large number of Gaussian likelihood evaluations required, lead to significant slowdowns in speech recognition processes.
Innovation Solution
Implementing a posterior-based feature with partial distance elimination, where log likelihoods for speech frames are calculated by sequentially adding dimensions, and the evaluation stops when the log likelihood is less than the lowest or a predefined threshold, reducing the number of dimensions needed for calculation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all dimensions are evaluated to calculate posterior-based features accurately, then recognition accuracy is improved, but computational cost increases significantly
Solution Approach 1:
The patent applies partial action by evaluating only a subset of dimensions rather than all dimensions. The system calculates partial distances using selected dimensions from the feature space, stopping when the partial distance exceeds a threshold or when sufficient accuracy is achieved. This reduces the number of Gaussian likelihood evaluations from potentially millions to a manageable number, directly addressing the contradiction between accuracy and computational cost.
Solution Approach 2:
The patent segments the dimension evaluation process into multiple stages. Instead of evaluating all dimensions simultaneously, the system processes dimensions in sequences, calculating partial distances incrementally. This segmentation allows the system to stop early when accuracy requirements are met, reducing overall computational burden while maintaining recognition accuracy.
2Measurement precision
If hundreds of thousands of Gaussian likelihood evaluations are performed, then posterior evaluation accuracy is improved, but speech recognition speed decreases
Solution Approach 1:
The system performs partial evaluations of Gaussian likelihoods by computing distances using only necessary dimensions. The partial distance calculation stops when the distance exceeds a threshold or when the posterior probability becomes sufficiently confident, avoiding unnecessary evaluations of remaining dimensions. This maintains accuracy while dramatically improving speech recognition speed.
Solution Approach 2:
The patent performs preliminary filtering by evaluating dimensions in an optimized sequence, placing more discriminative dimensions earlier in the evaluation order. This preliminary action allows the system to quickly eliminate unlikely hypotheses before performing full evaluations, improving speed without sacrificing accuracy.
Data Source
AI summary
A high-dimensional posterior-based feature with partial distance elimination may be utilized for speech recognition. The log likelihood values of a large number of Gaussians are needed to generate the high-dimensional posterior feature. Gaussians with very small log likelihoods are associated with zero posterior values. Log likelihoods for Gaussians for a speech frame may be evaluated with a partial distance elimination method. If the partial distance of a Gaussian is already too small, the Gaussian will have a zero posterior value. The partial distance may be calculated by sequentially adding individual dimensions in a group of dimensions. The partial distance elimination occurs when less than all of the dimensions in the group are sequentially added.