Speech Recognition Features Using Phonetic Unit Consistency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in accurately identifying spoken words due to inconsistencies between expected and detected phonetic units, lacking effective features to weigh the consistency and context-dependent patterns in audio signals.
Innovation Solution
The use of existence features, expectation features, and edit distance features, which are generated based on detected phonetic units and weighted according to learned correlations, are employed to determine the likelihood of a hypothesized word in an utterance, incorporating context-dependent and context-independent edit distance features to assess consistency and ordering information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech recognition models are used, then the system can process speech input, but the accuracy in identifying spoken words is insufficient due to inconsistencies between expected and detected phonetic units
Solution Approach 1:
The patent segments the speech recognition process into distinct feature components: existence features (whether a phonetic unit is present), expectation features (whether it matches the expected unit), and edit distance features (ordering relationships). This segmentation allows each aspect of phonetic unit detection to be independently evaluated and weighted, improving overall recognition accuracy despite inconsistencies in individual detections.
Solution Approach 2:
The patent transforms the speech recognition problem by changing the parameters used for evaluation. Instead of relying on a single model output, it introduces multiple binary and ordinal features (existence, expectation, edit distance) that capture different aspects of phonetic unit detection. These parameter changes enable the system to handle inconsistencies by weighing multiple evidence sources rather than relying on a single potentially unreliable indicator.
2Measurement precision
If multiple features are generated to improve recognition accuracy, then the differentiation between correct and incorrect detections improves, but the computational complexity increases
Solution Approach 1:
The patent divides the complex feature generation process into three distinct, manageable components: existence features (simple presence/absence detection), expectation features (comparison with expected units), and edit distance features (ordering relationships). Each component can be computed independently using relatively simple algorithms, avoiding the need for a single complex model while still achieving high recognition accuracy through their combination.
Solution Approach 2:
The patent introduces an intermediary layer of binary and ordinal features that mediate between the raw phonetic unit detections and the final recognition decision. These intermediate features (existence, expectation, edit distance) transform the raw detection data into a structured format that is easier for the recognition model to process, reducing the computational burden on the final decision-making stage while improving overall accuracy.
Data Source
AI summary
A computer-implemented speech recognition system described herein includes a receiver component that receives a plurality of detected units of an audio signal, wherein the audio signal comprises a speech utterance of an individual. A selector component selects a subset of the plurality of detected units that correspond to a particular time-span. A generator component generates at least one feature with respect to the particular time-span, wherein the at least one feature is one of an existence feature, an expectation feature, or an edit distance feature. Additionally, a statistical speech recognition model outputs at least one word that corresponds to the particular time-span based at least in part upon the at least one feature generated by the feature generator component.