雲间烟火Efficient Acoustic Front-End Processing for Tamil Speech Recognition using Modified GFCC Features
Fig.4. Spectrogram of the Pre-Processed Input Speech Signal for the
Word ―Poojiam‖ Uttered by Four Different Speakers.
One of the efficient techniques used for this purpose is windowing method. The most commonly used method is hamming window which has some limitations. Using hamming window, only one sample can be extracted in each segment. In general, taking a single sample from a signal does not provide an efficient estimate of its spectral properties and it does not help to achieve robust performance. The above problem can be solved by using the multi taper windowing method, where it is possible to take multiple independent spectral components from a same sample [19]. Multi Taper Windowing
In multi taper windowing, each data taper is multiplied element-wise to estimate the signal power at each frequency component. By optimizing a filter function, multiple windows can be derived which can effectively control the sidelobe and can prevent leakage of frequencies outside the bandwidth resolution. The
resultant windowed signal provides the statistically independent estimate of the underlying spectrum as each taper is pair-wise orthogonal to all the other tapers. The final spectrum is obtained by taking an average of all the tapered spectra. The experimental results have proved that the multi taper method provides small bias and low variance, as long as the PSD spectrum is flat and the frequency resolution is predetermined. The resultant signal obtained from multi taper windowing is then passed to the PSD for performing spectral analyzes. Yule-Walker AR Power Spectrum
