Sie sind auf Seite 1von 10

Computation Assignment -3

1) The duration of the window = 8000 * .03 = 240 samples.


Pre-emphasis is done with transfer function (1 - 0.95z-1) with α = 0.95.

Original Spectrum
After Pre- emphasis

2) For this, the windowed and pre-emphasized signal is used to calculate the
autocorrelation matrix ( r(0), r(1), r(2), .... r(p) ) , p is taken as 16 so that
autocorrelation values needed for other poles will also be obtained. These are given
as input to the levinson function which will give the LPC coefficients for each number
of poles. Using these coefficients the transfer function and hence the magnitude
spectrum corresponding to each number of poles is calculated using
G
H(Z) = -------------------------------------------------------------------------
1 + a1z-1 + a2z-2 - ......................................................aNz-N

Where N denotes the number of poles (6, 8, 10, 12 or 16) and


a1, a2, a3.... and aN are the coefficients obtained from the levinson function
G is the gain associated with the filter which is calculated using
G = (Emin)^(0.5),
Emin = r(0) - ∑ak*r(k), 1<=k<=N, N is the number of poles
aks are the coefficients obtained from the levinson function and
r denotes the autocorrelation function with the lag as k.
It is clear that as the order of the filter increases, the harmonics start to be followed
better.
LPC Spectrum Plot for P = 6

LPC Spectrum Plot for P = 8


LPC Spectrum Plot for P = 10

LPC Spectrum Plot for P = 12


LPC Spectrum Plot for P = 16

Pole-Zero Plot for P = 6


Pole-Zero Plot for P = 10

3) Inverse filtering is done using the equation


E(Z) = S(Z)A(Z)
The pre-emphasized waveform is passed through the filter with transfer
function A(Z).
A(Z) = 1 + a1z-1 +a2z-2 - ......................................................aNz-N
S(Z) denotes the Z-transform of the windowed and pre-emphasized speech
signal.
Pitch period can be calculated from the residual error signal and it is is found
as 63 samples which comes to about 126.9 Hz.
Residual Error Signal

Magnitude Spectrum of Residual Error Signal


4) The sound /a/ can be generated by passing the impulse train with F o = 126.9 Hz.
So an impulse train p(n) = delta(n-63k) k = 1,2,3....39 (To obtain 300ms waveform).
This p(n) is used to excite the LPC filter with 10 poles found in question 2. The
resultant sound is quite similar to /a/.

Input Impulse Train


Synthesized Sound

Synthesized Sound (zoomed in)

Das könnte Ihnen auch gefallen