Dialect Identification from Assamese speech using prosodic features and a neuro fuzzy classifier

摘要

Dialect identification is the task of classifying speech on the basis of dialect, which comes under the Automatic Language Identification problem. In this work a neuro fuzzy classifier is used to identify dialect of speech from vowel sound. Vowel sounds occur in an acoustic speech signal more frequently and with higher energy. Therefore, prosodic feature of vowel sounds can be used to search dialect dependent characteristics. First four formants of vowel sounds are used to create acoustic phonetic feature vectors and classify the speech segment as belonging to different dialect of the language. Fuzzy based systems are able to track minute variations in an environment invested with uncertainty. Minute variation of feature vectors are well classified by the neuro fuzzy classifier (NFC), which combines fuzzy classification with the learning capability of neural network. Experimental results are shown on the classification of four dialects of Assamese language mostly spoken in the North Eastern part of India. A comparison of NFC with Feed Forward Neural Network (FFNN) based classifier is also done. It is observed that NFC provides around 23% improvement in the correct classification rate compared to FFNN for the best feature set, which proves the effectiveness of NFC for dialect identification.