ECHO CANCELLATION USING MCLT

Mayur Agrawal, Subodh Gawade, Jagrutee Gawande, K.T.Talele

Sardar Patel Institute of Technology
Mumbai, INDIA

Abstract—Modulated Complex Lapped Transform defined by p (n, k) = pc (n, k) − j ps (n, k), where j is an
(MCLT) is a complex extension of the modulated lapped imaginary number and
 
pcn, k =  hn cos[n + k +  ]
transform (MLT). MCLT is a particular kind of a 2x π
oversampled generalized DFT filter bank, whose real part
corresponds to MLT. This property is used for efficient  
psn, k =  hn sin[n + k +  ]
π
(2)
noise cancellation, noise reduction and coding. For a  
length-M MCLT, the direct transform algorithm is based Where h (n) is the window function. Thus, we can also write
on computing a length-2M fast Fourier transform (FFT) X (k) = Xc (k) − j Xs (k), with j=√−1
plus M butterfly-like stages, without data shuffling. That
reduces the number of operations and memory accesses Xck = ∑ 
 xnpcn, k (3)
needed to compute the MCLT.
We see that Xc (k), the cosine components of X (k), are the
Keywords-mclt MLT of x (n). The window h (n) can be designed in many
ways, and a common choice is,
I. INTRODUCTION
 !
The modulated complex lapped transform (MCLT) is a hn = − sin[n +  ] (4)
cosine-modulated filter bank that maps overlapping blocks 

of a real-valued signal into complex-valued blocks of

transform coefficients. Thus, the MCLT performs a
frequency decomposition that is similar to that obtained
with the commonly-used Discrete Fourier Transform (DFT)
filter bank. The main advantage of the MCLT over DFT
filter banks is that the MCLT allows for three simple
reconstruction formulas, using all coefficients, only the
cosine coefficients or only the sine coefficients. That can be
useful in audio enhancement and encoding systems; for
example a noise reduction module can be implemented in
the MCLT domain, and after enhancement only the cosine
coefficients need to be sent to an encoder. In fact, the
MCLT cosine coefficients comprise the modulated lapped
transform (MLT) coefficients for the same signal, and the
MLT is quite popular in audio coding [1]. In MLT, the
transform coefficients are real and so they do not explicitly
carry any phase information. The MCLT has been used in
applications where magnitude-phase decomposition is
needed, as well as interfacing to coding systems, such as in
audio watermarking, audio stream identification, acoustic
echo cancellation and microphone array beam forming.
II. MODULATED COMPLEX LAPPED TRANSFORM
A length-M MCLT of a length-2M signal block x (n) is
defined by [1]
Xk = ∑

 xn ∗ pn, k (1)

' ] = −
[)8+2& + 1 −
Where the frequency index k varies from 0 to M – 1, n is the
time index, and p (n, k) are the analysis basis functions,
The inverse MCLT can be computed via the reconstruction
formula,
Yn = βc ∑
$ Xckpcn, k + βs ∑$ Xskpsn, k (5)

Where the coefficients β c and β s can be chosen in several
ways:
(1) βc = 1 and βs = 0 (cosine only)
(2) βc = 0 and βs = 1 (sine only)
(3) βc = βs = ½
III. NEW FAST MCLT ALGORITHMS
One disadvantage of the fast MCLT algorithm is that it does
not use a Fast Fourier Transform (FFT) directly. It maps the
MCLT into DCT-IVs, which can be computed via length-M/2
complex FFTs, but through a process that requires data
shuffling. Data shuffling means additional data buffer reads
and writes, which can take a considerable fraction of the total
processing time, especially in general-purpose processors,
where access to data memory is relatively slow (compared to
access to internal CPU registers). We now describe a new
MCLT algorithm that uses a single length-2M real FFT, with
no additional data shuffling. We start by assuming that M is
even, and that the window function h (n) is the common sine
window, which can be written as a sum of two complex
exponentials.
 ! (
ℎ& = − sin[& +
)8+−2& − 1] (6)
Where WM(r) is the common notation for the complex
exponential used in Fourier transforms, namely

( π.
WM(r)= exp  
'
On further calculation we get,

X(k) = jV(k) + V(k+1) (7)

Where,
V(k) = c(k) * U(k) (8)
c(k) = W8(2k + 1) * W4M(k)

/0 =  ∗ ∑ '
2 1& ∗ )2+0& (9)
'
We recognize U (k) as a normalized2 2M-point FFT of the
input block x (n). Then, it is clear from (7) and (8) that the Figure 2: Flow graph of the new fast inverse MCLT algorithm. The
input MCLT coefficients are combined to form a new set of
MCLT coefficients X(k) can be obtained by first computing coefficients; those are then sent to an inverse real FFT to generate the
the FFT of the input vector to obtain U(k) and then output signal. Arrows with hollow heads mean multiplication by –j,
performing the additional operations with the factors c(k), as and the shaded boxes represent the special mappings in.
shown in Figure 1. The coefficients c (k) satisfy |c (k)| = 1, 0.4
Input Signal

thus they are rotation operations [2]. That guarantees good

0.3
numerical properties of the algorithm, i.e. rounding errors
due to fixed- precision arithmetic have a minimal influence 0.2

on the final result.

0.1

-0.1

-0.2

-0.3
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
4
x 10

Figure 3: Signal corrupted by echo

Figure 1: Flowgraph of the fast MCLT algorithm

Similar calculations and derivation is done for inverse
MCLT. The flow graph for inverse MCLT is

Figure 4: Signal obtained after applying MCLT

using FFT
 Output Signal
0.25

0.2

0.15

0.1

0.05

-0.05

-0.1

-0.15

-0.2

-0.25
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
4
x 10

Figure 5: Noise Reduction by MCLT.

Thus we have implemented noise reducer using MCLT with

spectral subtraction [3]. The original speech was captured
near a noisy PC. We can see that the SNR is very low for
the above signal. In the result it shows that the SNR has
considerably increased & noise also reduced. More
importantly, the processed file has fewer artifacts than the
results obtained by using standard DFT filter banks.

IV. CONCLUSION
A new filter bank structure with complex-valued sub-bands,
which is referred to as MCLT - modulated complex lapped
transform is introduced. The MCLT is a simple extension of
the MLT, obtained by using both cosine and sine
modulations. This generates a subband structure that is 2x
oversampled in the frequency domain, because for every
new M real valued input samples the MCLT computes the
MCLT complex-valued frequency components. Using the
DCT-IV and DST-IV, we present fast algorithms for the
MCLT. They show that the MCLT has about twice the
complexity of the MLT
V. REFERENCES

[1] H. S. Malvar, “A modulated complex lapped transform

and its applications to audio processing,” Proc. Int. Conf.
Acoust., Speech, Signal Process., Phoenix, AZ, Mar.
1999, pp.1421–1424.
[2] H. S. Malvar, Signal Processing with Lapped
Transforms. Boston, MA: Artech House, 1992.