Dialect recognition from Telugu speech utterances using spectral and prosodic features

Abstract

The dialects of any language plays prominent role in speech processing applications. The identification of dialects is classifying the test speech samples to a particular dialect of spoken language of speaker. The crucial part of dialect recognition system is to discriminate the dialects of standard language because, there are lot of similarities between dialects of language. Dialect of particular language is not an unique feature as it is influenced by the nativity. The researchers of Automatic Speech Recognition encounter the limitation in analyzing the patterns of the speech that belongs to each dialect. This work identifies the dialects of Telugu language from unknown utterances of speech. Telugu Language is the one of the historical and popular language in Dravidian family. It contains mainly three dialects i.e. Rayalaseema, Telangana and Coastal Andhra. For identifying the dialects, we used different level of spectral and prosodic features of speech signal. The spectral features (MFCC, ∆ MFCC and ∆∆ MFCC) and prosodic features (Pitch and Loudness) of speech signals are considered to implement the system. As there is no standard database for Telugu dialects, dataset is created, to identify the dialects of Telugu language. To increase the accuracy in identifying the dialects, the spectral and prosodic features of speech utterances are combined. To identify the different dialects of Telugu language, two models for classification and experiments are carried out i.e. Gaussian Mixture Model and Hidden Markov Model models. The results are significant and accuracy of 88.40% for the GMM model and 86.95% for HMM model for Prosodic\u2009+\u2009MFCC feature is achievement