Spectral Features Based Spoken Dialect Identification for Punjabi Language

Dialect identification for speech based applications has a very bright future in a multi-lingual country like India. But due to the lack of resources like standard dialectal speech databases, it has not gained much importance. This paper focuses on identification of four dialects of eastern Punjab by utilizing the spectral features of speech. The dialectal speech corpus is also created for these four dialectal regions. For training and testing of the system KNN and MLP classifiers are implemented with different training and testing sets in the ratio of 70:30. The results showed that MLP clearly outperformed the KNN for different values of ‘K’ with the accuracy of 82.25% for MLP and 76.01%, 77.12% and 76.84% for KNN with K = 3, K = 5 and K = 7 respectively.