Confusing Phone Detection in Mandarin Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pronunciation training tools for Mandarin learners lack efficient automatic evaluation and high-level feedback, particularly in identifying and addressing confusing phones/syllables, and fail to consider users' nationalities and language backgrounds.
Innovation Solution
A method and system that analyzes non-native Mandarin speakers' pronunciation, identifies confusing phones/syllables, and provides real-time feedback by using a confusing-phone/syllable-embedded recognition net with Forced Alignment of Viterbi decoding, comparing segmented phones/syllables with a Mandarin acoustic model, and iteratively refining the recognition net to output incorrect pronunciation suggestions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If current pronunciation training tools partition and analyze Mandarin utterances to give scores, then automatic evaluation is provided, but confusing phones identification and high-level feedback are lacking
Solution Approach 1:
The system segments Mandarin utterances into individual phones and syllables, then analyzes each segment against acoustic models to identify confusing phones. This segmentation allows detailed phonetic analysis while maintaining automated evaluation, resolving the contradiction between automation and information loss.
Solution Approach 2:
The patent introduces an intermediary acoustic model comparison mechanism that acts as a mediator between the raw speech input and the final evaluation. This intermediary layer identifies confusing phones by comparing segmented phones with acoustic models, providing the missing high-level feedback while preserving automatic evaluation capabilities.
2Ease of operation
If pronunciation tools provide basic scoring feedback, then simple evaluation is achieved, but effective feedback considering users' nationalities and language backgrounds is not provided
Solution Approach 1:
The system applies local quality by providing differentiated feedback based on users' specific language backgrounds and nationalities. Instead of uniform scoring, the acoustic model comparison identifies confusing phones specific to each user's native language interference patterns, enabling personalized feedback while maintaining simple operation through automated detection.
3Measurement precision
If iterative refinement of recognition net is performed to improve pronunciation detection accuracy, then detection precision is improved, but processing time and system complexity increase
Solution Approach 1:
The patent applies preliminary action by pre-building the confusing-phone/syllable-embedded recognition net with known confusing phone patterns before actual pronunciation detection. This preliminary preparation enables rapid iterative refinement during detection without requiring complex real-time adjustments, improving detection accuracy while controlling system complexity through pre-computed acoustic models.
4Reliability
If confusing phones are identified and added to recognition net iteratively, then comprehensive confusing phone detection is achieved, but processing steps and time increase
Solution Approach 1:
The system implements continuous useful action through iterative refinement where each detection cycle builds upon previous results. Confusing phones identified in one iteration are immediately added to the recognition net for subsequent iterations, creating a continuous improvement process that achieves comprehensive detection efficiently by eliminating redundant analysis of already-identified confusing phones.
Data Source
AI summary
A method of generating and detecting confusing phones/syllables is disclosed. The method includes a generating stage and a detecting stage. The generating stage includes: (a) input a Mandarin utterance; (b) partition the Mandarin utterance into segmented phones/syllables and generate the most likely route in a recognition net via Forced Alignment of Viterbi decoding; (c) compare the segmented phones/syllables with a Mandarin acoustic model; (d) determine whether a confusing phone/syllable exists; (e) add the confusing phone/syllable into the recognition net and repeat step (b), (c), and (d) when the confusing phone/syllable exists; (f) stop and output all generated confusing phones/syllables to a confusing phone/syllable file when a confusing phone/syllable does not exist. The detecting stage includes: (g) input a spoken sentence; (h) align the spoken sentence with the recognition net; (i) determine the most likely route of the spoken sentence; and (j) compare the most likely route of the spoken sentence with the target route of the spoken sentence to detect pronunciation error and give high-level pronunciation suggestions.