Description
Consider the speech signal in “machali.wav”, sampled at 8 kHz. Consider the following signal
segments in the final word “pani”: (1) /a/ (first half); (2) /n/; (3) /I/; and (4) /s/ in the word “uska”.
Use PRAAT to extract the above segments to separate .wav files for further analyses as below.
1. Compute the narrowband spectrum using a Hamming window of duration = 30 ms before and
after pre-emphasis.
2. Using a 30 ms Hamming window centered in the pre-emphasized waveform:
(a) Compute the autocorrelation coefficients required for LPC calculation at various p =
4,6,8,10,12,20. Use the Levinson algorithm to compute the LP coefficients from the
autocorrelation coefficients. Show the pole-zero plots of the estimated all-pole filter for p=6,10.
(b) Compute the gain and plot the LPC spectrum magnitude (i.e. the dB magnitude frequency
response of the estimated all-pole filter) for each order “p”. Superimpose each plot on the
narrowband dB magnitude spectrum of part 1. Comment on the characteristics of the spectra.
(c) Plot error signal energy (i.e. square of gain) vs p.
3. Based on the 10th-order LPCs, carry out the inverse filtering of one of the vowel segments and
of the unvoiced sound. Obtain the residual error signal. Can you measure the pitch period of the
voiced sound from the residual waveform? Observe the magnitude spectrum of the residual
signal.
4. Next, we wish to resynthesize the phone sounds from the parameters of the source-filter model
obtained above for p=10: pitch, gain, LP coefficients. Use the LP filter estimated above, and an
ideal impulse train input as source excitation (for the voiced sounds). Carry out de-emphasis.
For the unvoiced sound, use a white noise signal as source excitation. Set the duration of the
synthesized sound to be 300 ms at 8 kHz sampling frequency, and view/listen to your created
sound.
Make a single document presenting your method with relevant code fragments, results and discussion.
