This paper presents a transformation-based approach to robust modeling of accented speech based on articulatory attributes for non-native speech recognition. Firstly, a two-stage verification method is used to extract speech segments from the speech input with non-native accent. Secondly, acoustic models of accented speech are transformed from normal models using linear transformation functions selected from a decision tree to deal with the problem of data sparseness. Thirdly, a discrimination function is applied to filter out the models with low recognition discriminability. Experimental results show that the inclusion of acoustic models of accented speech can eliminate recognition degradation in ASR due to non-native accents and the final ASR system can outperform the standard ASR system in recognizing non-native speech.
展开▼
