Optimally Sorting Elements for Speech Recognition Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for optimizing the coverage of elements in speech recognition systems, particularly during supervised and unsupervised adaptation, fail to efficiently minimize the number of training commands needed to achieve maximum phonetic coverage, leading to potential over-training and suboptimal adaptation.
Innovation Solution
A method that generates an optimally sorted list of elements by creating a weighted singleton histogram table, computing popularity scores for each singleton, and using these scores to prioritize elements for optimal coverage, thereby minimizing the number of commands required for training and preventing over-training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional speech recognition training uses all available commands, then comprehensive phonetic coverage is achieved, but the number of training commands increases leading to over-training and reduced efficiency
Solution Approach 1:
The patent segments the command list into prioritized subsets based on singleton phonetic content analysis. By dividing commands into priority groups (high, medium, low) based on their unique phonetic contributions, the system trains on essential segments first, achieving comprehensive coverage with fewer commands and preventing over-training on redundant examples
Solution Approach 2:
The patent changes the parameter of command selection from uniform random selection to priority-based selection weighted by singleton coverage metrics. By calculating and applying priority scores that reflect each command's unique phonetic value, the system transforms the training process to focus on high-value commands, reducing total training commands needed while maintaining phonetic coverage
2Adaptability or versatility
If unsupervised adaptation continuously processes all commands, then the system adapts to user voice, but the risk of over-training increases and adaptation efficiency decreases
Solution Approach 1:
The patent performs preliminary analysis of command singleton content before unsupervised adaptation begins, pre-calculating priority scores and identifying the most valuable command subsets. This preliminary action allows the adaptation process to immediately focus on high-priority commands, improving efficiency while maintaining adaptability
Solution Approach 2:
The patent applies partial action by selecting only the necessary subset of commands for unsupervised adaptation rather than processing all commands. By adapting on a prioritized subset that provides sufficient phonetic coverage, the system achieves effective voice adaptation with reduced computational load and lower over-training risk
3Reliability
If the speech engine uses a non-optimized command list for training, then all commands are covered, but the number of training iterations required increases and convergence is slower
Solution Approach 1:
The patent replaces the mechanical approach of sequential or random command processing with an intelligent priority-based selection system. By substituting uniform sampling with singleton-weighted priority sampling, the system efficiently identifies and processes the most informative commands first, reducing training iterations and time while maintaining comprehensive coverage
Data Source
AI summary
A method provides a non-optimized list of elements, with some of the elements having multiple terms. A table of sub-elements is generated from the elements list, with each sub-element having one term only and with a number of times a sub-element appears in the elements list being weighted in the sub-elements table. A weighted singleton histogram table is generated using a singleton dictionary, and a total popularity score of each singleton is computed from the sub-elements table. For each element from the elements list, an elements score is generated based on the total popularity score of each singleton within the element. An optimally sorted list of the elements list is generated based on the elements scores.