A

SPEECHENHANCEMENT METHOD KALMAN FILTERING

BASED UPON

K. T. Paliwal and Anjan Basu

Computer Systemsand Communication Group Tata Company of Important Research Homi Bhabha Road, Bombay 400005, India

SUBJECTIVE

In paper, problem this kind of the of s p e e c h n l a d c at the m electronic n big t e the moment corrupted sign only talk available is perfect for processingisconsidered. For this, theKalmanfiltering and compared the with Wiener m e t h o d is studied Its performanceisfoundtobesignifilteringmethod. f i actually c a n capital t le t t etrhn b sumado a ha elizabeth Wiener blocking method. A delayed-Kalman filtering method is as well proposed t h i actually c we m s r to v etsh s p e e c at the n they would a n c e m elizabeth np electronic r farreneheit o r m a n c e they would e h t of K a l m a and f we l to e ur f u r capital t h electronic r.

filter is designed for each Wiener short-time speech portion (duration= over 20 msec) by using a least-squares treatment. Though the non-stationary Wiener filter isoptimum for agivensegmentina does not least-squares-error perception, it take advantage of the knowledge presentation about production process. In the present paper, we all propose Kalman filtering technique which allows to get the nonstationarity of talk and, as well, exploits talk productionmodel. Wealsoshowthata delayed variation of the same filtering offers further more though improvement, the computat'ional complexity remains to be identical.

We. INTRODUCTION

a large number of In situations of practical fascination, the conversation signal gets corrupted eleven. KALMAN FILTRATION FOR TALK ENHANCEMENT by the addition of white noise. Existence of noise affects the intelligibility of your. Mathematical Formulation speech. The is the interaction between a pilot and an air traffic control Speechbe may represented by simply an structure, where conversation is usually degraded by autoregressive (AR) process is that the addiion of engine noises. such In essentialy the output of an all-pole situations, it truly is desirable enhance the to linear system powered by white noise quality and intelligibility of speech. In sequence. Thus presentation signal by k-th time automatic talk and audio recognition quick, s(k), is definitely givenby: devices if a presentation enhancement scheme is incorporatedinapreprocessing stage, s(k)=als(k-l)+.... +aps(k-p)+u(k) recognition becomes less difficult and more.... (1) reliable. Presentation enhancement as well plays a significant in position speech coding A little observation of the formula applications. (1) reveals it can be displayed by the state-space model asshown below. The condition addressed in our paper my spouse and i s to improve speech the moment only the damaged speech signal is available for processing. large A number of meth. ods have been completely reported in the literature [l] for presentation enhancement. standing The Wienerfilteringmethodisone of the... I(2) important presentation enhancement strategies. Since is speech non-stationary in nature, stationary Wiener filter will not perf u r ~ nvery very well. Theref ore, methods based on short-time power-spectrum have been proposed. Recently, Paliwal [ 2 ] offers proposed a non-stationary Wiener ilterirlg f method for talk enhancement, the place that the or

X(k)=

4

X(k-l)+ G u(k)

(3)

exactly where X(k), @ and G are state vector, express transion matrix and input matrix, correspondingly. are as These defined uses:

XT(k)=[s(k-p+l),..., s(k-l), ~(k)]#@@#@!!.... (4)

h

Queen,

=/

'---+-----I

I

We

I

We

I

one particular 0;

My spouse and i

(5)

I actually - a p... -al I (6)

!

covariancematrix is definitely P(klk). Likewise, of the one step forecasted estimate X(k) is X(kik-1) and affiliated error covariance matrix is P(kik-1). Using these note the Kalmanfilteringalgorithmcan be given recursive following the by simply relations: n

G sama dengan[0... 0 11

X(klk)=~(klk-l)cK(k)Cy(k)-H T(k: k-l)]

h

When only the noise corrupted signal y(k) is available, the observatlon method can be created in the next form: y(k)=s(k)+n(k) This equationcan be drafted in the type as follows: y(k)=H X(k)+n(k) where X(k) is the definedbyequation observation matrix (7:

X(k; kl)[email protected] 8(k-i Ik-l), with? Um ( F(k/k)=[I-K(k)H] P(kik-1)...