and after substitution of the composite components for I(Z) and H Z I . XfZI mav be exmessed as References 1 + rV(z)z-m - yV(2)Z-m 2 yZV(Z)Z-zm + 1+pz-- (10) Therefore the VSB induced terms cancel and the ghost is cancelled, resulting in X ( Z ) being the summation of the original video component, with a smaller double length ghost compared to the original. By simple rearrangement X(Z)may he expressed as 1 3 4 5 + 2pz-- + yzz-zm 6 The double length ghost may be cancelled by employing a biquadratic filter with an inverse response to the distortion component of X(z)as in Fig. 1, resulting in a deghosted signal. The deghost tilter may be extended to handle many ghosts by setting all the delay elements to cancel the longest ghost. The correct delay for each ghost may be realised by placing the relative tap coefficients closer to the input of the delay. This has the advantage that, for n ghosts, only n taps are required, unlike an FIR or transversal filter. Although the proposed deghoster requires four tapped line stores and seven multipliers, this is considered to be of little consequence with modem technology. 7 8 CICIORA, w., SGRIGNOLI, G., and THOMAS, w.: ‘A tutorial on ghost cancelling in television systems’, IEEE Trans., 1919, CE25, pp. 944 GARDINER, P.N., BROCKHURST, D.M., and CURTIS, J.s.: ‘Ghost cancellation for 625-line systems’. IBC94, 1994, IEE Conf. Pub., DD. 613418 KIM.K., OH,J., LEE,M., HWANG,H., and S0NG.D.: ‘A new ghost cancellation system for Korean GCR’, IEEE Trans., 1994, BC-40, pp. 132-140 E D W ~B.,~ ~~R ~R ,Y A., , WSTE, N., and GREENBERG, c.: ‘A singlechip video ghost canceller’, IEEE J. Solid-State Circuits, 1993, SSC-28, pp. 379-383 CHAO,T.S.: ‘Multi-path equalization for NTSC video by using digital IIR filter’,IEEE Trans., 1988, CE-34, pp. 268-278 CASE, D.R., and PARDOE, B.H.: ‘Measurement of multi-path effects in video sienals’. Fourth Int. Conf. on Television Measurements, 1991, IEE Conf. Pub. 335, pp. 63-68 BLUESTEIN, L I : ‘Envelope detection of vestigial sideband signals with application to television reception’, IEE Trans., 1975, CE-21, pp. 369-379 SHERRATT, R S., and PARDOE, B H : ‘The COIIcept Of a new video detector for ghosted IF signals’, to be published in Int. J. Electron., 1995 __ Idle-signal casting multiple access with collision resolution (ICMA/CR) protocol for wireless LAN 300 - U B.C. Kim, J.S. Jang and C.K. Un h 2 3 1 4 Results: Fig. 2 shows results (8-bit-sampled at 13.5 Msampleis) for both types of IIR deghoster discussed by applying an ideal detected VSB IF. Trace 1 shows the detected in-phase video pedestal line, trace 2 the detected quadrature signal, trace 3 the previous IIR deghoster output and trace 4 the new deghoster output. Both the video ghosting and VSB quadrature induced distortion can be seen on the detected signal traces 1 and 2. Residual VSB ghost effects may he seen on trace 3. Apart from quantisation noise, trace 4 deDicts no sien of ghosting or VSB induced distortion. A full compiement of ghost carrier phase and relative amplitudes yielded similar results. - - - Conclusions: The detection of the television I F VSB induced quadrature distortion of ghost carriers was shown to cause degradation in the operation of simple IIR ghost cancellers. However, by suitable processing of the detected quadrature signal, the VSB induced ghost distortion was cancelled from the in-phase detected signal before deghosting takes place, thus improving the performance of simpler IIR deghosters, which are intended to be placed within a commercial television receiver. 0 IEE 1995 I3 February 1995 Electronics Letters Online No: I9950467 R.S. Sherratt and B.H. Pardoe (Telford Research Institute of Acoustics, C/O Department of Electronic Engineering, Maxwell Building, University of Salford, Salford, M5 4 WT, United Kingdom) 698 Indexing terms: Multiaccess systems, Protocols, Pakcet radio networks, Local area networks The authors propose an idle-signal casting multiple access with collision resolution (ICMNCR) protocol for wireless LAN. This protocol focuses on efficient collision resolution by adopting the binary-tree protocol. By using the collision resolution algorithm, the retransmission process can be controlled so that collisions are resolved more efficiently using nearly immediate feedback information, thereby increasing the throughput-delayperformance of the ICMNCD protocol. Introduction: Recently, there has been much interest in the design of packet-radio local-area networks (LANs) for the third-generation wireless information network. As a result, to share a finite radio spectrum efficiently, researchers have studied many contention-based multiple-access schemes that adopt a variant of the carrier sense multiple access (CSMA) protocol. One of them is the idle-signal casting multiple access with collision detection (ICMN CD) scheme. According to the ICMNCD protocol, the central base station (BS) detects collision among several upward packets and broadcasts stop signals to improve the upward channel efficiency [I, 21. In the previous ICMNCD protocol, the major concern was on collision detection, but not on collision resolution. That is, packets receiving stop signals are rescheduled for retransmission at some later time, hoping that the collision problem will be solved in the future. It is well known that the binary-tree protocol is a stable and effective method for resolving collisions in the contentionbased access protocol. So far, however, there have been no studies on the performance characteristics of the collision resolution algorithm based on the binary-tree protocol in the ICMNCD protocol. In this Letter we propose an idle-signal casting multiple access with collision resolution (ICMNCR) protocol by adopting the binary-tree protocol 131. This protocol can resolve collisions effcciently as soon as they occur since nearly immediate feedback control signals available in local environments can be used. We describe the protocol operation, analyse its throughput-delay performance, and present the results which show the effectiveness of the protocol in collision resolution. ELECTRONICS LETTERS 27th April 1995 Vol. 31 No. 9 Protocol operation: We assume that N mobile users communicate with a central BS and that each user can be in one of two states: backlogged or thinking. In the thinking state, a user generates a new packet in a minislot with probability e,and then remains in the backlogged state until it completes successful transmission. All packets are assumed to have a fmed size equal to T. The BS broadcasts idle (I), busy (B) and stop (S) control signals over a separate channel to indicate transmission of no packet, one packet and more than one packet on the uplink, respectively. The BS can detect packet collisionswhen a failure occurs to synchronise to the fixed preamble pattern of received packets. We define the packet detection time D as the delay between the start of packet transmission at a terminal and its detection at another terminal by way of the central BS. (l/2)~+~j=,PI,~S~l/2~]/[l-(112)x-1] if k 2 3. For simplicity of analysis, we assume that all packets involved in collision during the kth CRI which starts with a collision among n packets at time t are successfully transmitted uniformly over [t + 3, f + B J as shown in Fig. 2. We denote E as the number of users who generate a new packet from the beginning of a CRI which starts with a collision among n packets to the end of CRIend pattern, providing that they are involved in collision at the beginning of a CRI. Also, we denote Q as those if they are not involved in collision. Then, we have P{I" = k } = I ; n n n: n d @ collision c idle i f collision / success @ success x %+,#ai p{On = k } - kn t du S" collision resolution interval ( C U I ) = J + 8, m Fig. 1 Timing chart of ICMA/CR protocol Let a collision resolution interval (CRI) be a time interval starting with the original collision (if any) and ending when this collision is resolved. At the beginning of a CRI, all users who have generated packets initialise the counter value r to zero and transmit packets, which initiate the zeroth cycle as shown in Fig. 1. All users involved in the original collision know when they can transmit their packets according to feedback information after packet detection time D and their counter value. This means that packets can he transmitted only I,, I, or I, time later after one has tried transmission according to the I, B or S control signal. Here we define a cycle as a time interval between successive packet transmission trials. Then, I,, 1, and 1, may be written as 1, = D + 1, 1, = D + T + 1, and I, = 2(D + I), respectively. We let r, be the counter value at the beginning of the tth cycle and x , the feedback information corresponding to the tth cycle. Also, we let x , = S, B and I represent collision, success and idle in the tth cycle, respectively. According to the received feedback x f , the updating process of the counter value obeys the following rule: If X~ = S and r, = 0, then r,+, = 0 or 1 with probability (1/2), and if x , = S and r, > 0, then r,,, = r, + 1. On the other hand, i f x , = B or I, then rr+l= r, 1. At the beginning of the tth cycle, only users with r, = 0 can transmit their packets. The BS broadcasts CRIend patterns(AA..AAI), where the number of As in the sequence is D, repeatedly on the downward control stream after detecting the end of the current CRI until new packets arrive. Users who have generated a new packet are not allowed to transmit until they hear CRIendpattern and initiate a new CRI only after hearing the last control information of CRIendpattern, which is I. 0;: = N,Ck{(Bn+ll)U}k{l-((Bn+ll)U}N-* P," = {(1-c/n)(Bn-Sn)+l1}o 1 5 c 5 downlink 1 (l-P:-,) ,az,---,ax 11 We define N"]as the initial number of packets involved in the original collision at the start of the kth CRI and select a Markov point at the beginning of a CRI. Then, a sequence of random variables {NW z = 1,2,...I constitutes a Markov chain, and the first-step transition probability from {N')= i } to {N'+l]= j ) is expressed as The steady-state distribution for the Markov chain {N'):t = 1,2,...I is given by N N i=O j=O Let U,be channel utilisation and Dp be the delay of a tagged packet, namely, the difference between its arrival time at the system and the time it is transmitted successfully. Then: U, = E:=, kskT TIBO+E;='=, nk(& +IT) ~ t t I &CRI e GSa n packets Fig. 2 Analytical model for approximate analysis 0.0 0.2 04 0.6 0.8 stationary throughput.packet /packet time 1.0 1116111 Fig. 3 Average stationary throughput-delay performance ( N = 10, T = 100) b. Performance analysis: We define as the length (in minislots) Of a CRI given that it starts with a collision among n packets and S. as the time interval between the original collision, if any, and the start of successful packet transmission. Let B. = 4 6 6 and S. = 43.1. Then, Bo = [ D + l ] / [ 1 4 4 ) ( D + ' ) B, ~ , = D+T+1 and B. = [1,+(1/2)"11,+~j=l"'.C~~1/2)"1]/[1~1/2)"1] if n 5 2. Also, SI = 0, S2 = [[2+(1/2)2]D+2+(1/2)~/[1-(1/2)] and S, = [[2+(1/2)*]0+ 2 + ELECTRONICS LEl7ERS 27th April 1995 Vol. 31 ~ 0 analysis simulation (ICMNCR) 0 simulation (ICMNCD) Results and dixussion: Fig. 3 shows the throughput-delay performance for error-free channels when D = 4 and 8, N = 10 and T = 100 (minislots). We compare the results with those obtained by No. 9 699 computer simulation of the ICMNCD protocol, in which the mean backoff period of each unsuccessful transmission is assumed to be 100 minislots. We can see that the proposed ICMNCR protocol can reduce the backoff period of unsuccessful transmission and resolve collision efficiently using the binary-tree protocol, and thus it yields higher throughput performance compared with the ICMNCD protocol. This improvement is more evident for short D and under light load because these conditions make CRI short in the ICMNCR protocol. 0 IEE 1995 24 February 1995 Electronics Letters Online No: 19950471 B.C. Kim, J.S. Jang and C.K. Un (Communications Research Laboratory, Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, 373-1 Kusong-Dong. Yusong-Gu. Taejon 305-701. Korea) References MURASE, A., and IMAMURA, K.: 'Idle-signal casting multiple access with collision detection (ICMA-CD) for land mobile radio', IEEE Trans., 1987, VT-36, pp. 45-50 2 LEE, I.H., and UN,c.K.: 'Performance analysis of 1-persistant ICMN CD with variable collision detection delay', Electron. Lett., 1993, 29, pp. 1562-1563 3 ROM, R., and SIDI, M.: 'Multiple access protocols: Performance and analysis' (Springer-Verlag, Inc., New York, 1990), pp. 107-147 1 To obtain a good separation of the projected data, the difference between the two means must be large relative to some measure of the standard deviation for each class. Then, the Fisher linear discriminant is defined as the linear function w'x for which the following criterion function is maximised: J(w) is a distance measure for the sample means on the projected line, where S , = SI + S , is the within-class scatter matrix (covariance) and S , = (m, - m2)(ml - m,)' is the between-class scatter matrix. This expression is well known in mathematical physics as the generalised Rayleigh quotient [3]. It can be derived that the value of w which maximises J(w) is w = S,-l(m, - m,). Substituting w into eqn. 1, we obtain J(w) = (m, - m,)'S,-l(m, - m2). Here this sample mean measure is introduced as a distance, called the RQ measure, to evaluate the difference between two speakers' statistics. The covariance and the sample mean are both involved in the computation, but the measure only pays attention to the sample mean difference. (ii) Absolute deviation measure: Using the distance between two real symmetric matrices, we can construct measures in many different ways. One possibility is to perform an operation on the matrix absolute difference Cd = IC, - C,l. Let d,,(C,, C,) he the trace of matrix C,: k where hk is the kth eigenvalue of C,. A group of asymmetric measures was discussed in [I], where all distances are based on operations of the matrix ratio C,C,-'. Let h; be the ktb eigenvalue of C,C,-l; then an absolute deviation measure is defined as Nonparametric distance measures of speaker verification K.T. Ng, H. Li, J.-P. Haton and J. Su which performs the same operation as our trace measure but on C,C2-' - I rather than C , - C,. Noting Indexing terms: Speaker recognition, Signal processing nonparametric technique for speaker recognition and verification is proposed. The proposed distance measures differ from existing distance measures by their symmetry. These measures have been evaluated on a text-dependent database, achieving a 99.6% verification rate for 200 French speakers. In addition, it is shown that the covariance carries more speaker information than the sample mean. A Introduction: In practical speaker verification systems, it is not realistic to ask for a large amount of speaker training data. Therefore, parametric modelling usually leads to poor estimation. Some nonparametric models have been applied to text-independent cases [l] with excellent results; however, the measures used [l] are asymmetric distance measures. Some earlier work [l, 21 has also pointed out that short term dynamics of speech carries important speaker information. In this Letter, the statistics of covariance, correlation and mean on individual speech frames, along with short-term speech dynamics, concatenating m adjacent frames of speech data into a long frame, are used and shown to be important speaker discriminative characteristics. IC1 - Czl = I(ClC,' - 1)CZl = I(C,C,' - 1)ClI (4) the asymmetry lies in the fact that C, # C,. The measure of eqn. 2, termed TRl, evaluates distances between correlation matrices. Replacing correlation with covariance matrices, we obtain another measure, TR2. For the three kinds of measures, the RQ measure focuses more on the difference of sample mean, while TR1 focuses on correlation only, and TR2 on both correlation and sample mean. The measure of eqn. 3, referred to as TR3 and another asymmetric measure, the sphericity measure [l] (SM), are also assessed in the experiments for comparison. Verification scoring: Speaker verification can be considered as a traditional two-category Bayes decision problem of 'true' or 'false'. It is known that a normalised distance outperforms an unnormalised distance by using a collection of speakers as background speakers, termed the cohort. Two issues should thus be addressed one is the number of background speakers, another is Nonparametric measures: Let an N-frame observation of an utterthe selection of the speakers. The speaker-specific cohort for a ance be X = (xI,...,x",...,x N } , where x. is in K-dimensional space. speaker is usually selected by choosing some speakers nearest to 4,___, him in the training population. This kind of cohort is shown to be We then have a sequence of speech dynamics D = {d,,.__, dN-ol+l},where d. is a concatenation of m frames {x",...,xn+,,+,}. good at discriminating acoustically similar speakers but poor for Therefore, there exists a sample mean m = (l/N)&xn, a correlation acoustically dissimilar speakers. Therefore, the cohort size should matrix C = (l/N)Xnxnx; and a covariance matrix S = (UN)& be large enough to cover a distribution of features over the popu(x, - m)(x, - m)' for an observation sequence or a short-term lation. However, increasing the cohort size may introduce more dynamics sequence. computation. Here, a new cohort selection is proposed. For each speaker, (i) Rayleigh quotient measure: A traditional pattern classification who could be considered as the cohort leader, we choose the mproblem is to classify data from two classes XIand X,. One possinearest and m-farthest speakers as the cohort members. The mbility is to project a K-dimensional space onto a line, y = w'x, so nearest speakers are obtained from the cohort selection proposed that we can estimate the scattering of two classes in one dimenin previous papers [2]. In designing a Bayes classifier, we introduce sion. a set of two events { A l , R,} for speaker i. A, means that the claim ELECTRONICS LETERS 700 -~ - 27th April 1995 Vol. 31 No. 9

1/--страниц