Every individual has some unique speaking style and this variation influences their speech characteristics. Speakers’ native dialect is one of the major factors influencing their speech characteristics that influence the performance of automatic speech recognition system (ASR). In this paper, we describe a method to identify Hindi dialects and examine the contribution of different acoustic-phonetic features for the purpose. Mel frequency cepstral coefficients (MFCC), Perceptual linear prediction coefficients (PLP) and PLP derived from Mel-scale filter bank (MF-PLP) have been extracted as spectral features from the spoken utterances. They are further used to measure the capability of Auto-associative neural networks (AANN) for capturing non-linear relation specific to information from spectral features. Prosodic features are for capturing long - range features. Based on these features efficiency of AANN is measured to model intrinsic characteristics of speech features due to dialects.
展开▼
