Документ взят из кэша поисковой машины. Адрес оригинального документа : http://imaging.cs.msu.ru/pub/2010.DSPA.Tkachenko_Lukin.Denoising.ru.pdf
Дата изменения: Tue Apr 6 23:00:00 2010
Дата индексирования: Mon Oct 1 19:49:58 2012
Кодировка:

_____________________________________________________________________________________________

The same procedure is applied to a question asked by user. Then some kind of intersection of two sets is obtained. The first set is previously obtained set for text documents and the second set is obtained by applying morphological analysis and hypotheses generation procedures to text representation of a question. The resulting answer set is obtained by extraction of words that are consistent with the question word from that intersection and sorting it in decreasing order of ranks. For now proposed method is restricted to provide only one word answers without taking into account the context in which the words are used. In future those restrictions will be removed. The proposed method of generating and ranking of hypotheses was implemented with a number of restrictions and tested on a database of news articles (over 3500 documents). Studies of the developed system showed positive results and proved its suitability to the task. Now the main goal of research is to develop algorithms of deductive, inductive and abductive logical inference. Those algorithms will help to answer such questions, the answers to which are not contained in the indexed documents in explicit form.

. ., . 1. . : x[t ] , : ( y[t] . . . . [1]. . ( . .« , . . (Non-Local Means) . (Multi-Resolution FFT: MR FFT), .. ). STFT : ( . 2. : ,f . 3. « » (noise footprint) , . , NOISE[f, t] ,t STFT : ) [3]. (STFT), : , » , (Short Time Fourier Transform: STFT) . ). [2]. , , , , , . , , x[t] , ) .. ,,

..

y[t ] noise[t ] ,

t

-

-

-

. . -

( 2. 1.

-

FOOTPRINT [ f ]
,

1 K

K t1

NOISE[ f , t ] ,

,K

Y [ f , t ] max{ X [ f , t ] k W [ f , t ], 0}

(

)

_____________________________________________________________________________________________

223
Digital signal processing and its applications

_____________________________________________________________________________________________

X[f, t]

W[f, t] , k . STFT STFT. , , . ( . , . .

, Y[f, t] -

4.

, ,

«

»

). , .

, -

3. , , , . ( . 1).

. 1. , MR FFT STFT, : , , . STFT ( ,
r

[4, 5].

,

-

(

),

, . .
r

). , ,
N

.

r

ai

r

.

a

S

r

i1

i

N

N i1

a

2 i

, ( , 4. ), ( ,

STFT c ). . , STFT. , 12
r

. STFT
r

. . . w:

1/16
= (S ize-1) w , 1, 1 1

M ixFFT + CurrentFFT

M ixFFT -C urrentFFT +1 CurrentFFT *InvOverlap ,

224

_____________________________________________________________________________________________ 12Proceedings of the 12-th International Conference

_____________________________________________________________________________________________

MixFFT STFT ).

STFT , InvOverlap STFT. .

, (samples). Size 64

Size

, CurrentFFT STFT (1/4, r STFT, -

, . 1/4 . 5. , , [6]. 1 : . [7]. 11 13.10 ,
r

,

. , -

,

-

11

. .

-

93

13.15

93 .

.

16.10

. .

-

1. 6.

16.70 .

-

, STFT . , . . « » 20092013 . .

STFT . , -

-

1. David Gibson. The Art of Mixing: A Visual Guide to Recording, Engineering, and Production. REV, 2005. ISBN13: 9781931140454. pp. 52-53, 162 p. 2. Gilbert Strang, Truong Nguyen. Wavelets and Filter Banks. Wellesley-Cambridge Press, 1996. ISBN: 0961408871, 9780961408879. pp. 103-105, 490 p. 3. Saeed V. Vaseghi. Advanced Digital Signal Processing and Noise Reduction. John Wiley & Sons Ltd, 2000. ISBN: 0471626929. pp. 5-8, 466 p. 4. Alexey Lukin, Jeremy Todd. Adaptive Time-Frequency Resolution for Analysis and Processing of Audio. AES'06, October 2006, 10 p. 5. Karin Dressler. Sinusoidal Extraction Using an Efficient Implementation of a Multi-Resolution FFT. DAFx'06, September 2006, 6 p. 6. Antoni Buades, Bartomeu Coll, Jean-Michel Morel. A non-local algorithm for image denoising. CVPR'05, Vol. 2, June 2005, pp 60-65. 7. : http://makseq.com/makseq

_____________________________________________________________________________________________

225
Digital signal processing and its applications

_____________________________________________________________________________________________

A MULTIRESOLUTION SPECTRAL SUBTRACTION ALGORITHM FOR NOISE SUPPRESSION IN AUDIO SIGNALS Tkachenko M., Lukin A. Moscow State University, Faculty of Computational Mathematics and Cybernetics, Laboratory of Mathematical Methods of Image Processing Noise is an undesirable signal appearing during transmission or measurement of another clean signal. There are two categories of noise, by spectral properties: stationary (the one that does not change in time) and non-stationary. Additive noise is summed with the clean signal y[t] and does not depend on it: x[t] = y[t] + noise[t], where t is time and x[t] is the observed signal. Constant hiss from a microphone or an amplifier, electric hum are the examples of additive stationary noise. In this paper, we propose a method for suppression of additive stationary noise using a multiresolution STFT (short-time Fourier transform). This approach is able to improve quality of many audio processing algorithms (such as noise reduction) by adaptively varying STFT time-frequency resolution based on local properties of the signal: estimates of spectrogram sparsity. The simple method of spectral subtraction (SS) is widely used for reduction of additive stationary noises. It works with a fixed STFT window size. Long windows provide good frequency resolution and achieve accurate separation of closely spaced noise and signal harmonics. However long windows sizes also lead to poor time resolution and increase smearing of transients (sharp attacks or fast changes in the signal). On the other hand, shortwindow STFT processing is inefficient in terms of frequency resolution and possible depth of noise suppression. A filter bank with a fixed window size cannot provide good frequency and time resolution simultaneously. It is required to select best resolution for each local part of the signal during the processing. We propose using a multiresolution STFT to solve this problem. The algorithm of MR STFT consists of three parts: 1. Calculation of several copies of the signal processed with different STFT window sizes; 2. Estimation of smearing (or sparsity) for local spectrogram areas for each resolution; 3. Mixing of the resulting spectrograms based on spectrogram sparsity estimates. The effect of this adaptive algorithm is selection of high frequency resolution for tonal signals and selection of high time resolution for transients. A noise reduction system based on MR STFT has been created. Several modifications of spectral subtraction rules have been implemented, including a highly effective method of Non-Local Means for smoothing of a "musical noise" artifact. During algorithm testing we have found optimal range of STFT sizes for the MT STFT frameworks, sizes of spectrogram mixing areas and other key parameters. We have compared the performance of a simple spectral subtraction (SS) with several STFT window sizes, SS based on MR STFT (SS MR STFT), and SS with a Non-Local Means based on MR STFT. It has been shown that SS MR STFT produces results significantly different with a simple SS: after processing with a simple SS noise power decreased by 13.00 dB, while with SS MR FFT by 16.10 dB. The application of a Non-Local Means smoothing has removed musical noise and further improved overall processing quality.

.. ... , , . . , , -

s (t )

fS , s (n t ) ,
.

n

,

n ... 2, 1, 0,1, 2, ... , t 1 / f S s ( n) ,

-

n 0,1, ... , N 1 ,

~ S ( m)

nN1 n0

s(n) w(n)e

2 jmn f t

a ( m)

jb(m) ,

226

_____________________________________________________________________________________________ 12Proceedings of the 12-th International Conference