SEMI-SYLLABLE UNITS FOR ROBUST TEXT INDEPENDENT SPEAKER IDENFICATION

Abstract In this study, Robust text-independent speaker identification is investigated. Syllable and semisyllableboundaries are automatically detected in Farsi continuous speech utterances using short-term energy contour and discrete wavelet transform (DWT). While the entire syllable is considered as the unit for prosody, wavelet entropy coefficients are emerged from two overlapping semi-syllables reflecting consonant/vowel (CV) and (or) vowel/consonant (VC) transitions distinctly. Long-term prosodic features i.e. rational syllable nuclei duration, mean energy; pitch frequency and, four formants in addition to concatenated coefficients of wavelet entropy in depth four are extracted as the feature vector. Classification is performed by the feed-forward perceptron neural network (FFPNN) with two hidden layers. The experiments conducted on Farsi speech dataset (FarsDat) using proposed method confirm improvement in speaker identification accuracy in different signal to noise atios compared with conventional methods

Keywords:

Text-Independent Speaker Identification , Discrete Wavelet Transform , Semi-Syllables , Prosody , Wavelet Entropy