Accent Classification in Speech Using the i-Vector Framework in Language Proficiency Platforms

Speech accentology is an important subject of language proficiency platforms as it allows the correct evaluation of pronunciation and regionalism. Proper recognition of accents in speakers assists in customization of learning experiences and enhancement of automated evaluations of language. The current accent recognition techniques tend to lack robustness when analyzing short speech segments and they vary among speakers thereby decreasing the classification accuracy. Its i-vector based traditional methods are effective in the recognition of speakers but less effective in the extra accent-specific features in short utterances. To overcome these shortcomings, this paper presents a Deep Segmental i-Vector Approach (DSiVA), which is a combination of a segmental feature extraction and deep neural network modeling. DSiVA successfully represents local accent properties through the segmentation of speech into meaningful units and the creation of i-vectors of these units, and the deep network combines this information to classification by a better means. This structure increases the resistance to speaker variation and brief utterances, which offers a more accurate accent classification system. The suggested DSiVA technique works with several language proficiency datasets to check the efficiency of the technique in differentiating accents between various speakers. According to the results of the experiment, DSiVA is more effective than traditional i -vector and baseline deep learning models because the first one is more accurate and consistent in recognising accents. The results suggest that it can support adaptive language learning systems and computerbased speech assessment systems.

Hali tarjima qilinmagan