Speech Recognition Alternate Selection via Classifier Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in providing efficient error correction methods for users, as they often present lengthy lists of alternates, which can be burdensome and reduce user experience, especially on mobile devices where screen space is limited.
Innovation Solution
A classifier-based approach is employed to selectively display a subset of alternates to users, using machine learning techniques such as logistic regression to estimate the probability of an alternate being useful, thereby reducing the average length of alternates lists without compromising the number of correctable errors, and ranking them based on classifier outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all speech recognition alternates are displayed to users, then the number of correctable errors is maximized, but the length of alternates lists increases making the interface burdensome
Solution Approach 1:
The patent extracts only the most relevant alternates from the complete set of speech recognition hypotheses. A classifier identifies and extracts a subset of alternates that have the highest probability of being useful to the user, displaying only these selected alternates rather than all possible alternatives. This extraction principle resolves the contradiction by maintaining error correction capability while reducing list length.
Solution Approach 2:
The patent changes the parameter of alternate selection from displaying all hypotheses to displaying a filtered subset based on classifier scores. The classifier computes a relevance score for each alternate and uses this parameter to determine which alternates to display, transforming the display list from a complete set to an optimized subset that balances correctness and brevity.
2Length of moving object
If a classifier is used to select alternates, then the average length of alternates lists is reduced, but the system complexity increases
Solution Approach 1:
The patent introduces a classifier as an intermediary component between speech recognition and the user interface. This mediator evaluates alternates and determines which ones to display, acting as a bridge that translates the complete set of hypotheses into a optimized display list. The classifier serves as an intermediary that automates the selection process, reducing the need for complex manual curation while maintaining list quality.
3Reliability
If more alternates are displayed, then users can correct more errors, but screen space is consumed especially on mobile devices
Solution Approach 1:
The patent extracts only the essential alternates needed for effective error correction, removing redundant alternatives that would consume screen space. By extracting a focused subset of high-value alternates, the system maintains error correction capability while minimizing the area occupied on the display, particularly benefiting mobile devices with limited screen real estate.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting alternates in speech recognition. In some implementations, data is received that indicates multiple speech recognition hypotheses for an utterance. Based on the multiple speech recognition hypotheses, multiple alternates for a particular portion of a transcription of the utterance are identified. For each of the identified alternates, one or more features scores are determined, the features scores are input to a trained classifier, and an output is received from the classifier. A subset of the identified alternates is selected, based on the classifier outputs, to provide for display. Data indicating the selected subset of the alternates is provided for display.