Multi-level Subword Parse Table for Accented Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional speech recognition models fail to accurately recognize accented speech due to inconsistent pronunciations, leading to high error rates, as they are optimized for native speech and lack effective contextual modeling.
Innovation Solution
A multi-level subword parse table is employed to model words at different subword levels (syllable, position-specific cluster, and phone) to determine probabilities of pronunciation variations in accented speech, combining these probabilities to form a robust pronunciation model for improved recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional speech recognition models are used, then native speech recognition works adequately, but accented speech recognition produces high error rates
Solution Approach 1:
The patent segments words into multiple subword levels (phones, position-specific clusters, syllables) and creates separate parse tables for each level. This segmentation allows the system to model pronunciation variations at different granularities, improving accented speech recognition by capturing patterns that single-level models miss.
Solution Approach 2:
The patent combines probability information from multiple subword levels (phone-level, PSC-level, syllable-level) into a composite pronunciation model. This composite approach integrates diverse linguistic information sources to create a more robust model that handles accented speech better than any single level could alone.
2Reliability
If pronunciation rewriting is used to adapt the dictionary, then accented speech recognition improves, but large amounts of input data are required
Solution Approach 1:
By segmenting the pronunciation modeling task into multiple subword levels, the system can leverage existing linguistic knowledge and data at each level rather than requiring massive amounts of accented speech data for complete retraining. Each level contributes independently to the overall model.
Solution Approach 2:
The multi-level parse table approach creates a universal framework that can handle various accents and pronunciation variations without requiring accent-specific training data for each case. The same structural framework adapts to different accents by combining probabilities from its multiple levels.
3Ease of manufacture
If phone-level rewriting rules are used, then pronunciation adaptation is achieved, but contextual information at higher levels is not utilized
Solution Approach 1:
The patent segments the linguistic information into hierarchical levels (phones, PSCs, syllables) and creates separate parse tables for each. This segmentation preserves contextual information at each level while maintaining the simplicity of phone-level rules, allowing the system to use both detailed phonetic information and higher-level contextual patterns.
Solution Approach 2:
The patent merges probability information from multiple subword levels into a unified pronunciation model. By combining phone-level, PSC-level, and syllable-level probabilities, the system recovers contextual information that would be lost in phone-level-only approaches while maintaining computational feasibility.
Data Source
AI summary
Techniques are described for training a speech recognition model for accented speech. A subword parse table is employed that models mispronunciations at multiple subword levels, such as the syllable, position-specific cluster, and/or phone levels. Mispronunciation probability data is then generated at each level based on inputted training data, such as phone-level annotated transcripts of accented speech. Data from different levels of the subword parse table may then be combined to determine the accented speech model. Mispronunciation probability data at each subword level is based at least in part on context at that level. In some embodiments, phone-level annotated transcripts are generated using a semi-supervised method.