HW 06: LINEAR PREDICTION AND THE CEPSTRUM

  1. Select a vowel for a low-pitched male speaker, and a high-pitched female speaker. Sample this signal at 8 kHz and 16 kHz. Center a 30 msec analysis window about the center of this vowel. Plot the linear prediction spectrum using the autocorrelation method for orders 2, 4, 8, 16, and 32 for both samples frequencies and both speakers. Interpret the results. (Hint: Also compute the LP error energy.)

  2. Generate the linear prediction error signal for the 8th order analysis performed above for the 8 kHz sample frequency. Plot the waveform and the residual in a time-aligned display. What conclusions can you draw about the location of the peaks in the LP residual relative to the waveform?

  3. For the male speaker, separate the excitation from the vocal tract using cepstral processing. Plot the time domain signals and the quefrency domain signals that result from this processing, following the format presented in class.

  4. For this same utterance, compare the DFT-derived cepstral coefficients to the LP-derived cepstral coefficients.