Fundamental frequency estimation using the dyadic wavelet transform

This thesis explores the use of the Dyadic Wavelet Transform for the analysis of continuous speech. In this context, after a brief discussion on how pitch measurement in real-time is important for the analysis of continuous speech, the Dyadic Wavelet Transform is discussed. This thesis also discusses how to estimate the word boundary using the Dyadic Wavelet Transform. A detailed description on the design of an analyzing filter is presented here. The performance of a Dyadic Wavelet transform based pitch estimation algorithm is evaluated using a non-orthogonal wavelet (cubic spline) and compared to the Autocorrelation based pitch estimation algorithm. The results from processing sentence consisting of eight words show that the Dyadic Wavelet Transform algorithm has better performance than an Autocorrelation based algorithm. All the results are presented in the form of percentage relative errors. For the estimation of the word boundary the Dyadic Wavelet Transform based algorithm has shown very impressive results with less than 2% relative error. Computationally, the Dyadic Wavelet Transform based algorithm is simpler than the classical methods due to its matrix formulation. For a signal sampled at 16 kHz a 32 msec window is selected in order to get at least two peaks that correspond to pitch. For a 32 msec window, sampled signals are taken at every 40 sample. In this way, one can measure the minimum pitch period of the speech signal. A minimum 50% overlapping is required for avoiding edge effect due to inadequate wavelet coefficients.;Various recommendations are also made for improving the performance of the Dyadic Wavelet Transform based algorithm. A hybrid continuous speech recognition system using Dyadic Wavelet Transform and conventional method is proposed.