Automatic Spotting of Vowels, Nasals and Approximants from Speech Signals

摘要

Automatic speech recognition involves methodologies for translation of spoken language into text. An important problem that needs to be solved for the success of speech recognition is the accurate detection of phonemes. In this paper, a two stage system for spotting the boundaries of vowels, nasals and approximants in Malayalam speech signal is proposed. In the first stage, speech signal is classified into six broad phoneme classes using an Artificial Neural Network based broad phoneme classifier. Classifier with nine features has limited accuracy for detecting vowel, nasal and approximant boundaries. So features like difference of spectral spread, spectral centroid, envelope variance, energy ratio, difference in formant frequencies are added to the classifier. With these additional features, a major improvement in classifier accuracy is achieved. In the second stage, a frequency domain parameter named spectral peak frequency is suggested for accurate verification of nasals. Sonorant and nonsyllabic features are used for verifying approximants and syllabic feature is used for locating vowels.