Speech Recognition Model Segmentation for Non-Native Accent Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face challenges in recognizing non-native languages due to the need for extensive training materials and resources, particularly in mixed-language scenarios like Chinese-accented English, where accent mismatches and pronunciation variations among native speakers complicate the collection and quality of training data.
Innovation Solution
A speech recognition system that generates and updates speech-units based on a speech recognition model with native and non-native components, using modules like the 'speech-analysis and unit-expansion module' to select optimal components and update the model without requiring extensive retraining, allowing for recognition of non-native words without significantly impacting native language performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional speech recognition systems use single-language models for multiple-language recognition, then device complexity is reduced, but recognition precision deteriorates due to accent mismatches and pronunciation variations
Solution Approach 1:
The patent segments the speech recognition model into separate single-language models (first language model and second language model) instead of using a single multi-language model. Each model is trained on native speaker data for its specific language, allowing the system to maintain high recognition precision for each language while avoiding the complexity of building and managing a unified multi-language model.
Solution Approach 2:
The patent creates a universal speech recognition system that can handle multiple languages by combining separate single-language models. The system achieves multi-language capability through the integration of specialized models, where each model performs its specific language recognition function optimally, and the system as a whole provides universal multi-language support.
2Measurement precision
If separate speech recognition models are trained for each language, then recognition precision is improved, but resources and efforts required increase significantly
Solution Approach 1:
The patent uses copying by leveraging existing high-quality native speaker training materials for each language to create specialized single-language models. Instead of collecting new training data for each language, the system copies and adapts proven training approaches from single-language systems, reducing the overall resource requirement while maintaining high recognition precision.
3Measurement precision
If separate speech recognition models are trained for each language, then recognition precision is improved, but device complexity increases
Solution Approach 1:
The patent implements a dynamic model selection mechanism that automatically chooses the appropriate single-language model based on the detected input language. The system dynamically switches between the first language model and second language model, providing a simple interface for users while managing multiple specialized models in the background, thus reducing the perceived complexity for end users.
4Manufacturing precision
If non-native language training materials are collected from native speakers of that language, then speech-units quality is improved, but ease of operation deteriorates due to accent mismatches and pronunciation variations
Solution Approach 1:
The patent copies high-quality speech data from native speakers of the target language to create training materials for non-native speech recognition. By copying and adapting proven training datasets from single-language systems, the system achieves high speech-unit quality without the difficulty of collecting and verifying non-native speaker data, as the training materials are replicated from established native speaker corpora.
Data Source
AI summary
A speech recognition system and method thereof, a vocabulary establishing method and a computer program product are provided. The speech recognition method includes: storing a speech recognition model including speech-units and basic components of acoustic models, wherein each of the speech-units includes at least one state and each state corresponds to one of the basic components of acoustic models; receiving first and second speech signals; obtaining a speech-unit sequence of a native/non-native vocabulary from a speech-analysis and unit-expansion module; recognizing the first speech signal according to the speech recognition model and the speech-unit sequence of the native/non-native vocabulary and further outputting a recognition result; and selecting an optimal component from the basic components of acoustic models according to the speech recognition model, the second speech signal, and the word corresponding to the second speech signal, and further updating the speech-units according to the best basic component of acoustic model.