A System for Automatic Regional Accent Classification
Guntur Radha Krishna, Ramakrishnan Krishnan, Vinay Kumar Mittal
Abstract
Automated identification of a speaker's native language is carried out using non-native English speech data spoken by native speakers of languages Kannada, Tamil, Telugu. Dataset consists of speech in text-dependent and text-independent modes. Experiments are conducted using Mel-frequency cepstral coefficients (MFCCs), and classifiers Gaussian Mixture Models (GMM), GMM with Universal Background Model (UBM) and i-vector. A prototype system for Automated Classification of Speakers based on Regional Accent (ACSRA) is developed. Best results are obtained using the i-vector based classifier. Accuracy at 93.9% is obtained for native regional language identification (NLI), using the text-independent speech data. It is observed that distinguishing the native language is better using the text-independent data than with text-dependent data. Further analysis indicated better classification accuracy for female non-native speakers (87%), than for male speakers (80%). It is found that identifying English speakers of Kannada is easier than that of Telugu or Tamil languages.