Speech Recognition Disambiguation via Multi-Stage ML Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face challenges in accurately processing spoken utterances due to ambiguity in disambiguation, where multiple hypotheses are generated with similar semantic interpretations, leading to user frustration and incorrect command execution.
Innovation Solution
A two-stage configuration using machine learning models to determine whether disambiguation is necessary and which hypotheses to present to the user, where the first model assesses confidence scores to decide on execution or disambiguation, and the second model selects hypotheses for user selection, with additional processing to remove similar results before disambiguation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all generated hypotheses are presented to the user for disambiguation, then the accuracy of command execution is improved, but the user experience deteriorates due to frustration from seeing multiple similar options
Solution Approach 1:
The patent extracts and removes hypotheses that have similar semantic interpretations from the set of hypotheses presented to the user. By identifying and eliminating redundant options that would confuse the user, the system maintains high accuracy in command execution while improving user experience by showing only distinct, meaningful choices.
Solution Approach 2:
The patent changes the parameter of hypothesis selection by introducing a similarity threshold parameter. Hypotheses are filtered based on their semantic similarity scores, and only those exceeding the threshold are removed. This parameter-driven approach allows the system to dynamically adjust which hypotheses are presented, balancing accuracy and user experience.
2Productivity
If a single hypothesis is selected for execution without disambiguation, then the system operates efficiently, but the reliability of command execution deteriorates due to potential incorrect selections
Solution Approach 1:
The patent performs preliminary disambiguation processing before final hypothesis selection and execution. By evaluating semantic similarity and confidence scores in advance, the system determines whether disambiguation is necessary, ensuring reliable command execution while avoiding unnecessary delays for clear cases.
Solution Approach 2:
The patent implements a feedback mechanism where the system evaluates the confidence scores and semantic similarity of hypotheses, then uses this information to decide whether to present disambiguation options to the user or proceed with automatic selection. This feedback loop ensures reliable execution by adapting the disambiguation process to the specific characteristics of each set of hypotheses.
3Measurement precision
If disambiguation processing is performed for all hypotheses, then the accuracy of command execution is improved, but the processing time increases
Solution Approach 1:
The patent applies partial disambiguation processing by selectively evaluating only those hypotheses that meet certain confidence and similarity thresholds. Instead of performing exhaustive disambiguation on all hypotheses, the system focuses computational resources on the most ambiguous cases, reducing processing time while maintaining accuracy.
Solution Approach 2:
The patent uses parameter thresholds for confidence scores and semantic similarity to control the extent of disambiguation processing. By adjusting these parameters, the system can optimize the balance between processing time and accuracy, performing full disambiguation only when necessary based on the characteristics of the generated hypotheses.
Data Source
AI summary
Automatic speech recognition (ASR) processing including a two-stage configuration. After ASR processing of an incoming utterance where the ASR outputs an N-best list including multiple hypotheses, a first stage determines whether to execute a command associated with one of the hypotheses or whether to output some of the hypotheses of the N-best list for disambiguation. A second stage determines what hypotheses should be included in the disambiguation choices. A first machine learning model is used at the first stage and a second machine learning model is used at the second stage. The multi-stage configuration allows for reduced speech processing errors as well as a reduced number of utterances sent for disambiguation, which thus improves the user experience.