Scribd Logo
Upload

Welcome to Scribd!

  • 30 views

    l8 Ceps

    Uploaded by

    Ghanshyam V Rao
    The document discusses homomorphic signal processing and the cepstrum. It begins by defining the cepstrum as the inverse Fourier transform of the log magnitude spectrum. A homomorphic system is one where the principle of superposition is obeyed after a nonlinear transformation. Speech signals can be modeled as the convolution of excitation and vocal tract impulse responses. Taking the log and inverse Fourier transform separates these components in the complex cepstrum domain. Short-time cepstral analysis using the DFT approximates this. Homomorphic filtering can then select the desired component by liftering in the quefrency domain.

    Copyright:

    Attribution Non-Commercial (BY-NC)
    Download as PDF, TXT or read online from Scribd

    l8 Ceps

    Uploaded by

    Ghanshyam V Rao
    30 views1 page
    The document discusses homomorphic signal processing and the cepstrum. It begins by defining the cepstrum as the inverse Fourier transform of the log magnitude spectrum. A homomorphic system is one where the principle of superposition is obeyed after a nonlinear transformation. Speech signals can be modeled as the convolution of excitation and vocal tract impulse responses. Taking the log and inverse Fourier transform separates these components in the complex cepstrum domain. Short-time cepstral analysis using the DFT approximates this. Homomorphic filtering can then select the desired component by liftering in the quefrency domain.

    Original Title

    l8_ceps

    Copyright

    © Attribution Non-Commercial (BY-NC)

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document discusses homomorphic signal processing and the cepstrum. It begins by defining the cepstrum as the inverse Fourier transform of the log magnitude spectrum. A homomorphic system is one where the principle of superposition is obeyed after a nonlinear transformation. Speech signals can be modeled as the convolution of excitation and vocal tract impulse responses. Taking the log and inverse Fourier transform separates these components in the complex cepstrum domain. Short-time cepstral analysis using the DFT approximates this. Homomorphic filtering can then select the desired component by liftering in the quefrency domain.

    Copyright:

    Attribution Non-Commercial (BY-NC)
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    30 views1 page

    l8 Ceps

    Uploaded by

    Ghanshyam V Rao
    The document discusses homomorphic signal processing and the cepstrum. It begins by defining the cepstrum as the inverse Fourier transform of the log magnitude spectrum. A homomorphic system is one where the principle of superposition is obeyed after a nonlinear transformation. Speech signals can be modeled as the convolution of excitation and vocal tract impulse responses. Taking the log and inverse Fourier transform separates these components in the complex cepstrum domain. Short-time cepstral analysis using the DFT approximates this. Homomorphic filtering can then select the desired component by liftering in the quefrency domain.

    Copyright:

    Attribution Non-Commercial (BY-NC)
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    You are on page 1of 1

    The Cepstrum and Homomorphic Processing

    Introduction The Quefrency Alanysis of Time Series for Echoes In general, we nd ourselves operating on the frequency side in ways customary on the time side and vice versa

    Cepstrum Denition The cepstrum of a signal x[n] is dened by 1 j jn c[n] = log |X (e )|e d 2 where X (ej ) is the DTFT of x[n]. From (1), we can see that c[n] is the inverse DTFT of the logrithm of the magnitude spectrum log |X (ej )|. Complex Cepstrum The complex cepstrum is dened by 1 j jn [n] = log X (e )e d x 2 Homomorphic Systems Oppenheim dened classes of non-linear systems based on a generalized principle of superposition. Such systems are called homomorphic systems. The Principle of Superposition For a linear system L(), the principle of superposition is obeyed. That is, L(x1[n] + x2[n]) = Lx1[n] + Lx2[n] (3) L(cx[n]) = cLx[n] as shown in the following gure
    + L() x[n] +

    (1)

    (2)

    y[n] = L(x[n])

    Figure: Conventional Linear System

    Homomorphic Systems For a homomorphic system H(), the generalized principle of superposition is obeyed. The actually representation depends on the operation of interest. In the case of convolution, we have y [n] = H(x1[n] x2[n]) = Hx1[n] Lx2[n] = y1[n] y2[n] as shown in the following gure
    H() x1 [n] x2 [n]

    (4)

    y[n] = H(x1 [n]) H(x2 [n])

    Figure: A Homomorphic System for Convolution

    Canonical Homomorphic System for Convolution A homomorphic system (for convolution) is equivalent to a cascade of three homomorphic systems depicted as follows
    D () x[n] x1 [n] x2 [n] + x[n] + L() + y [n] y1 [n] + y2 [n] +
    1 D ()

    y[n] y1 [n] y2 [n]

    x1 [n] = x2 [n]

    Figure: Canonical System for a Homomorphic System for Convolution


    1 D() is called the characteristic system for convolution, while D () is called the inverse characteristic system for convolution. Both are (normally) xed systems. The linear system L() is to be designed for special applications. The input-output relations of the subsystems are as follows. x [n] = D(x1[n] x2[n]) = Dx1[n] + Dx2[n] = x1[n] + x2[n] y [n] = L(x1[n] + x2[n]) = Lx1[n] + Lx2[n] = y1[n] + y2[n] (5) 1 1 1 2[n]) = D y1[n] D y2[n] = y1[n] y2[n] y [n] = D (y1[n] + y

    Representation by DTFT A characteristic system for homomorphic deconvolution using DTFT is


    D () F () x[n] X(ej ) + F 1 () X(ej ) x[n]


    log()

    Figure: Characteristic System for Homomorphic System using DTFT

    We have

    X (e ) =
    n

    xn[n]e

    jn

    , (6)

    X (ej ) = log{X (ej )}, 1 j jn X (e )e d x [n] = 2 It is easy to show that z[n] = x[n] y [n] z [n] = x [n] + y [n]

    (7)

    Representation by z-Transform A similar characteristic system can be dened using z-transform instead of DTFT. See Figures 8.9 and 8.10. Computation of the Complex Cepstrum Rational z-Transforms A rational z-transform can be represented by where ak is a zero inside the unit circle, ck is a pole inside the unit circle, and 1 bk is a zero outside the unit circle. Taking the logarithm, we have
    Mo Mi 1 log |bk | k=1 Mo

    X (z) =

    Mi k=1(1

    ak z

    Ni k=1(1

    Mo k=1(1 ck z 1)

    1 1 bk z )

    (8)

    X (z) = log |A| + +


    k=1

    + log(z

    Mo Ni

    )+
    k=1

    log(1 ak z

    ) (9)

    log(1 bk z)
    k=1 n

    log(1 ck z 1)

    Using the power series, log(1 a) =


    n=1

    a , n

    |a| < 1,

    (10)

    we obtain the complex cepstrum 1 log |A| + Mo log |bk |, k=1 Ni n Mi n ck ak n n, x [n] = k=1 k=1 M o n bk n ,
    k=1

    n=0 n>0 n<0 (11)

    Homomorphic Analysis of Speech Signal

    Our basic assumption for speech production is based on convolution. Specically, for voiced speech, we assume s[n] = p[n] hV [n], where hV [n] = AV g[n] v [n] r [n]. (12) For unvoiced speech, we assume s[n] = u[n] hU [n], where hU [n] = AU v [n] r [n]. (13) In (12) and (13), g[n], v [n], and r [n] are the glottal pulse, the vocal tract impulse response, and the radiation load response respectively. p[n] is the quasi-periodic excitation, u[n] is the unvoiced excitation signal with unit variance, and AU , AV are the gains. Example of Voiced Speech Figure 8.13 shows the respective time-domain functions of g[n], v [n], r [n], p[n] for a model for sustained vowel /AE/. Figure 8.14 shows the pole-zero plots of the respective z-transforms. Figure 8.15 shows the log magnitudes of the respective DTFTs. Figure 8.17 shows the respective complex cepstra. Figure 8.18 shows the speech complex cepstrum and cepstrum. Example of Unvoiced Speech For unvoiced speech, there is no glottal pulse. The periodic excitation is replaced by a random noise function u[n]. The autocorrelation representation is thus modeled by ss [n] = vv [n] rr [n] The DTFT of (14) is ss (e ) =
    j j 2 j 2 2 AU |V (e )| |R(e )| , 2 AU [n]

    2 AU vv [n]

    rr [n].

    (14) (15) (16)

    so the complex cepstrum of the autocorrlation function is ss [n] = 2 log AU [n] + (v [n] + v [n]) + ([n] + [n]) r r Figure 8.20 shows an example for unvoiced /AE/. Short-time Cepstral Analysis To extract a window of speech of length L beginning at n, we dene a nite-length sequence xn[n] = w[n]s[n + n] 0
    L1

    0n L1 otherwise

    (17)

    The DTFT of xn[n] is the short-time Fourier transform of s[n] Xn(e ) =


    n=0 j

    xn[n]e

    j n

    (18)

    (6) is modied to be
    L1

    Xn(e ) =
    n=0

    xn[n]e

    j n

    , (19)

    Xn(ej ) = log{Xn(ej )}, 1 j jn xn[n] = Xn(e )e d 2

    Short-time Cepstral Analysis Based on DFT Using DFT, (19) is simplied (approximated) to
    L1

    Xn[k ] =
    n=0

    xn[n]e

    j 2 kn N

    0k N 1 (20)

    Xn[k ] = log{Xn[k ]}, 0 k N 1 [n] = 1 X [k ]ej 2 kn, 0 n N 1 N xn n N It can be shown that xn[n] = xn[n + rN], 0 n N 1
    r

    (21)

    Homomorphic Filtering of Natural Speech We assume that over the length of window, say L, the speech signal satises s[n] = e[n] h[n], (22) where e[n] is excitation and h[n] is the impulse response from glottis to lip. Furthermore, we assume that h[n] is short compared to L, so x[n] = w[n]s[n] = w[n](e[n] h[n]) ew [n] h[n], where ew [n] = w[n]e[n]. Voiced Speech Let e[n] be a unit impulse train of the form
    Nw 1

    (23)

    e[n] =
    k=0

    [n kNp ],

    (24)

    where Np is the pitch period (in samples), and Nw is the number of impulses in a window. The windowed excitation is
    Nw 1

    ew [n] = w[n]e[n] =
    k=0

    wNp [k ][n kNp ], 0 k Nw 1 otherwise

    (25)

    where w[kNp ] wNp [k ] = 0 Taking the DTFT of (25), we have


    Nw 1

    (26)

    Ew (e ) =
    k=0

    wNp [k ]e
    2 . Np

    jkNp

    = WNp (e

    jNp

    ),

    (27)

    which is periodic in with period

    It follows that the log spectrum (28)


    j j

    (ej ) = log HV (ej ) + log Ew (ej ) X

    has two components: a slow-varying log HV (e ) and a periodic log Ew (e ). Note that as log Ew (ej ) is also periodic, and the inverse DTFT produces an impulse train with period Np . Thus the two components in x [n] = hV [n] + ew [n] (29) can be distinguished. Using DFT The above procedure can be carried out by DFT as well. The system is depicted in Figure 8.31. Liftering The desired component of the input can be selected by a window function in quefrency, denoted by l[n]. This is also called liftering. For excitation signal, a high-pass ltering of the log spectrum is used, while for the impulse response, the low-pass ltering is used. Figure 8.32 shows an example.

    National Sun Yat-Sen University - Kaohsiung, Taiwan

    You might also like