Journal of ELECTRICAL ENGINEERING, VOL. 62, NO.

1, 2011, 16

01 01
02 02
03 LOW COMPLEX FORWARD ADAPTIVE LOSS COMPRESSION 03
04 04
05
ALGORITHM AND ITS APPLICATION IN SPEECH CODING 05
06 06
07 Jelena Nikolic Zoran Peric Dragan Antic 07
08 08
Aleksandra Jovanovic Dragan Denic
09 09
10 10
11 This paper proposes a low complex forward adaptive loss compression algorithm that works on the frame by frame basis.
11
12 Particularly, the algorithm we propose performs frame by frame analysis of the input speech signal, estimates and quantizes 12
13 the gain within the frames in order to enable the quantization by the forward adaptive piecewise linear optimal compandor. 13
14 In comparison to the solution designed according to the G.711 standard, our algorithm provides not only higher level of 14
the average signal to quantization noise ratio, but also performs a reduction of the PCM bit rate for about 1 bits/sample.
15 15
Moreover, the algorithm we propose completely satisfies the G.712 standard, since it provides overreaching the curve defined
16 by the G.712 standard in the whole of variance range. Accordingly, we can reasonably believe that our algorithm will find 16
17 its practical implementation in the high quality coding of signals, represented with less than 8 bits/sample, which as well as 17
18 speech signals follow Laplacian distribution and have the time varying variances. 18
19 K e y w o r d s: forward adaptive technique, Loss compression algorithm, piecewise linear optimal compandor 19
20 20
21 21
22 1 INTRODUCTION Take a notice of the fact that an additional reduction over 22
23 the PCM bit rate can be achieved by applying some of 23
24 Numerous research has been conducted during the re- the lossless compression algorithms, as it has been shown 24
25 cent years with the goal to develop a coding algorithm in [11], where the Ramalho G.711 lossless compression 25
26 that minimizes the bit rate in the digital representation 26
algorithm has been applied. The lossless compression al-
27 of a speech signal without a significant loss of the sig- 27
gorithms are usually applied after the loss compression of
28 nal quality in the process. Although a great number of 28
the signal has been performed. Therefore, it is important
29 speech coding algorithms has been developed [15], there 29
to research the suitable loss compression algorithm that
30 is still an indication of the reasonable need for continu- 30
provides the highest possible signal quality for the given
31 ation of the research in this field [6]. We decided to find 31
bit rate. After the suitable choice of the loss compression
32 the solution to the formulated problem by means of wave- 32
algorithm has been performed, any of the lossless com-
33 form coders, since they provide the highest level of speech 33
pression algorithms can be applied to provide a further
34 quality [1 ,68]. 34
bit rate reduction. Accordingly, in this paper we have de-
35 35
The simplest and the most commonly used waveform cided to focus on the field of loss compression algorithms.
36 36
coding algorithm, defined by the G.711 standard [9], pro-
37 Regarding the time varying characteristics of speech 37
vides high quality speech at 64 kb/s. Particularly, it pro-
38 signals, we directed our research towards the field of adap- 38
vides the conversion of 12 bits samples to 8-bit code by
39 tive coders [1, 2, 7, 8] that attempt to make the encoder- 39
using companded 8 bits/sample Pulse Code Modulation
40 decoder (coder or quantizer) designs adaptable to the 40
(PCM) [1, 69]. Up to now, so much work has been done
41 varying characteristics of the input signals. Furthermore, 41
in the field of loss compression algorithms in order to
42 provide not only an additional reduction over PCM bit since backward adaptation provides the SQN R within 42
43 rate, but also to provide the highest possible quality of 1 dB of forward adaptation, as well as since forward adap- 43
44 the digitized speech signal (measured by SQN R Sig- tation is less sensitive to transmission errors when com- 44
45 nal to Quantization Noise Ratio). The most significant pared to backward adaptation [8, 11], we have destined 45
46 result of such a research is the development of the adap- to perform our research in the field of forward adap- 46
47 tive predictive speech coders, which has brought the bit tive waveform coding algorithms. Regarding the small- 47
48 rate for high quality speech coding to 16 kb/s provid- est average complexity of scalar compandor model [13], 48
49 ing a reduction by a factor of four over the PCM bit rate our research has started with the forward adaptive scalar 49
50 [1, 4, 6, 10]. However, since it has already been shown that compandor. Particularly, we decided to provide an addi- 50
51 the complexity of the adaptive predictive speech coders tional complexity reduction by performing the lineariza- 51
52 can be considerably high [1, 4, 10], we were motivated to tion on the nonlinear forward adaptive scalar compan- 52
53 find the manner to reduce the PCM bit rate by means of dor. Accordingly, in this paper we propose the forward 53
54 low complex non-predictive waveform coding algorithm. adaptive piecewise linear optimal companding coding al- 54
55 55
56 Faculty of Electronic Engineering Nis, Aleksandra Medvedeva 14, 18000 Nis, Serbia, jelena.nikolic@elfak.ni.ac.rs 56
57 57
58 c 2010 FEI STU
ISSN 1335-3632 58
 2 J. Nikolic Z. Peric D. Antic A. Jovanovic D. Denic: LOW COMPLEX FORWARD ADAPTIVE LOSS COMPRESSION . . .

01 01
02 02
03 03
04 04
05 05
06 06
07 07
08 08
09 09
10 Fig. 1. Forward adaptive coding scheme with non adaptive scalar Fig. 2. Forward adaptive coding scheme with adaptive scalar quan- 10
11 quantizer tizer 11
12 12
13 13
gorithm that presents the information about the input of a buffer, an adaptive N -level piecewise linear optimal
14 14
signal with approximately 7 bits/sample. compandor (PLOC), the gain estimator and the Ng -level
15 15
We will discuss the performance (SQN R and the bit scalar quantizer for gain quantizing (SQgq ). Particularly,
16 16
rate) of the proposed algorithm, which has been deter- we have decided to implement the log-uniform scalar
17 17
mined in a wide variance range of the input speech sig- quantizer (for gain quantizing), since we have recently
18 18
nals. Additionally, experimental results of the real speech demonstrated that it could provide higher SQN R than
19 19
signal processing will be analyzed in order to practically the uniform scalar quantizer [15]. Moreover, we have des-
20 20
test the performance of the proposed coding algorithm. In tined to consider the piecewise linear optimal compandor,
21 21
order to point out the benefits of the proposed algorithm, since it has already been pointed out that although the
22 22
the achieved performance will be compared to the appro- smooth and differentiable nonlinear compressor charac-
23 teristics are convenient for mathematical manipulations, 23
priate one corresponding to the speech coding algorithm
24 there are problems of accurately implementing analog 24
defined by G.711 standard [9]. Finally, in order to point
25 nonlinearities [7]. Particularly, it has been ascertained 25
out that the proposed speech coding algorithm provides
26 that the solution to these problems provides todays tech- 26
a high quality speech coding, G.712 standard [14] that
27 nology which allows the implementation of uniform quan- 27
defines the smallest SQN R , which has to be achieved for
28 tizers with a piecewise linear compressor characteristics 28
a high quality transmission, will be used.
29 that can approximate the smooth compressor curve. 29
30 The design procedure of the proposed coding scheme 30
31 2 NOVEL FORWARD ADAPTIVE consists of the following steps: 31
32 CODING ALGORITHM 32
Step 1. Design of non adaptive (fixed) PLOC for the ref-
33 2 33
erence variance (ref = 1 ) is based on finding the seg-
34 Adaptive coding schemes or algorithms work on frame- 34
ment thresholds denoted by tj , j = L, L + 1, . . . , L.
35 by-frame basis, where a frame consists of a certain num- 35
Namely, a piecewise uniform quantizer is a quantizer
36 ber of samples [1, 5, 7, 8, 12]. The conceptual difference 36
whose amplitude range consists of several segments,
37 between the adaptive coding algorithms is based on the 37
each of which contains several quantization cells and
38 manner in which the adaptation is performed whether 38
output levels corresponding to a uniform quantizer [7].
39 it is performed forward, ie, from the input sequence or It is well known that by dividing the output region of the 39
40 backward, ie, from the coded output signal [5, 7, 8, 12]. N -level compandor to 2L equidistant regions, where L 40
41 Due to the reduced sensitivity of the forward adaptive is the number of segments (regions) in the first quadrant, 41
42 technique to transmission errors when compared to the the region of input signal is, according to the obtained 42
43 backward adaptive technique, as well as since it has been piecewise linear compressor characteristic, divided in 2L 43
44 demonstrated that the backward adaptive technique pro- non equal segments each of which contains n = N/2L 44
45 vides SQN R within 1 dB of the forward adaptive tech- quantization cells and output levels that correspond to 45
46 nique [8, 12], we have decided to focus our research on a uniform quantizer [7]. 46
47 the forward adaptive coding algorithms. Forward adapta- Step 2. Buffering of the input signal and the gain esti- 47
48 tion can be performed by normalizing the input sequence, mation Buffering frame after frame enables an esti- 48
49 further coding with non adaptive quantizer (coder or mation of the gain, defined as /ref , ie as a ratio of 49
50 encoder-decoder) and finally by performing the denor- the squared root of the frame variance and squared root 50
51 malization procedure with the same quantized value of of the reference variance [7] 51
52 the calculated gain that was used for normalizing (see 52
53 Figure 1) [1, 8]. The same effect can be achieved by us-
q
1 PM1 53
54 i=0 x2j+1 54
ing an adaptive quantizer with a code book obtained g=
M
p , (1)
55 by multiplying the code book of the aforementioned non 2
ref 55
56 adaptive quantizer with the quantized value of the esti- 56
57 mated gain [7 ,8]. Here we have destined to base our algo- where a frame consists of a certain number of samples 57
58 rithm on such a coding scheme (see Figure 2) consisting xj+i , i = 0, 1, . . . , M 1 . 58
 Journal of ELECTRICAL ENGINEERING 62, NO. 1, 2011 3

01 Ng levels, and then, the quantized gain obtained in such 01

02 a manner is used for adjusting the codebook of the for- 02
03 ward adaptive PLOC. Next, the encoding procedure is 03
04 performed for the current frame and the code word in- 04
05 dex I is obtained as the result of the encoding process. 05
06 Since the encoding procedure at the transmitting end of 06
07 the transmission system is normally followed by the de- 07
08 coding procedure at the receiving end of the transmission 08
09 system, the information about the quantized gain is nec- 09
10 essary for the decoder. Therefore, the side information S 10
11 is sent along with the codeword index I . After the side in- 11
12 formation is decoded, the decoding procedure that results 12
13 in the output of the forward adaptive PLOC quantizer is 13
14 provided. We believe that the proposed algorithm is very 14
15 simple and suitable for implementation and therefore easy 15
16 for practical use. In order to test the proposed algorithm, 16
17 we will consider its application in speech coding. 17
18 18
19 3 APPLICATION IN SPEECH CODING 19
20 20
21 The quality of a quantizer (coder) is usually measured 21
22 by the distortion of the resulting reproduction x in com- 22
23 parison to the original signal x. Observe that the distor- 23
24 tion introduced by the proposed coding solution [7] 24
25 25
26 L
X aj 2 26
27 D = D(, g, ref ) = 2 Pja , (4) 27
12
28 i=1 28
29 29
30 depends on the PLOC quantization step sizes aj 30
31 Fig. 3. Forward adaptive PLOC coding algorithm 31
32 taj (
g , ref ) taj1 ( g, ref ) 32
aj = aj (
g , ref = ,
33 n (5) 33
34
Step 3. Quantization of the estimated gain by using the j = L, L + 1, . . . , 1, 1, 2, . . . , L , 34
Ng -level log-uniform scalar quantizer
35 35
36 as well as on the probability that the input lies in the 36
37
lu j -th segment 37
20 log(
g = gk ) = 20 log min + (2k 1) ,
38 2 38
max (2)
39 20 log min Z ta
j (
g,ref ) 39
k = 1, . . . , Ng , lu = , Pja = Pja (, g, ref ) = p(x, )dx ,
40 Ng (6) 40
ta
j1 (g,ref )
41 41
where the variance range of the input signal in decibels j = L, L + 1, . . . , 1, 1, 2, . . . , L .
42 42
43 (20 log min , 20 log max ) is divided into Ng cells having 43
44 equal lengths lu . Let us now define the signal to quantization noise ratio 44
45 Step 4. Design of adaptive PLOC The finding of dependence on the signal variance in order to ascertain 45
whether G.712 standard can be satisfied with the pro-
46 decision thresholds of the adaptive PLOC, denoted by 46
47 taj , is enabled by multiplying the segment thresholds of posed coding algorithm [7, 8] 47
48 the non adaptive PLOC (obtained in the step 1) with 48
i2
49 the quantized gain g SQN R = 10 log . (7) 49
50 D(i ) 50
taj = taj g, ref = gtj ref , j = L, L + 1, . . . , L . (3)



51 51
52
Additionally, let us define both, the theoretical and the 52
53
experimental average signal to quantization noise ratio 53
According to the described design procedure and the SQN Ra , SQN Raex [8] in order to provide an instant com-
54 block diagram shown in Fig. 3, we can summarise that 54
parison of the theoretical and the experimental results
55 the novel coding algorithm firstly performs buffering of 55
56 the current frame, after which it performs an estimation k
56
57 of the gain for the considered frame. The estimated gain 1X i2 57
SQN Ra = 10 log , (8)
58 is further quantized by the log-uniform quantizer having k i=1 D(i ) 58
 4 J. Nikolic Z. Peric D. Antic A. Jovanovic D. Denic: LOW COMPLEX FORWARD ADAPTIVE LOSS COMPRESSION . . .

01 01
02 02
03 03
04 04
05 05
06 06
07 07
08 08
09 09
10 10
11 11
12 12
13 13
14 14
15 15
Fig. 4. Experimental results: Dependence of SQN Rex
a on the frame Fig. 5. Experimental results: Illustration of the SQN Rex
p vicis-
16 length M ( W = 10200 , N = 128 , Ng = 32 )
16
situde through the frames having length M = 33 ( W = 10200 ,
17 N = 128 , Ng = 32 , R = 7.15 bits/sample) 17
18 18
19 19
20 20
21 21
22 22
23 23
24 24
25 25
26 26
27 27
28 28
29 29
30 30
31 31
32 32
33 33
34 Fig. 6. Experimental results: Dependence of SQN Rex 34
a on the frame Fig. 7. Experimental results: Illustration of the SQN Rex
p vicis-
35 length M ( W = 10200 , N = 128 , Ng = 64 ) situde through the frames having length M = 49 ( W = 10200 , 35
36 N = 128 , Ng = 64 , R = 7.12 bits/sample) 36
37 37
38 PF PM 38
p=1 q=1 x2pq and taking into account the expression for the optimal
39 SQN Raex = 10 log PF PM , (9) compressor function that was derived for such a distribu-
39
(x a )2
ypq
40 p=1 q=1 pq 40
tion [7]
41 41
where k denotes the number of the particular variances x


42 1 exp c xmax 42
a
that are considered, while xpq and xpq = ypq denote the c(x)opt = xmax
, (12)
43 1 exp c 43
44 input samples and the outputs of the adaptive PLOC 44
45 quantizer, respectively. In order to provide a more de- c = C/3 , C = xmax /ref , we have derived the following 45
tailed analysis of the performance achieved by performing experession for the positive segment thresholds of the
46 46
an experiment on the real speech signal, we have decided nonadaptive PLOC
47 47
48 to define the average signal to quantization noise ratio xmax L 48
tj = ln , j = 0, 1, . . . , L . (13)
49 within the each of F frames (each having M samples) c L j(1 exp(c )) 49
50 Note that the segment thresholds are symetric to the cor- 50
51 1 PM
x2pq responding counterparts. Finally, combining the set of 51
M q=1
52 SQN Rpex = 10 log 1
PM , p = 1, . . . , F . equations (1), (2), (3), (13) with the result from [16], 52
(x a )2
ypq
53 M q=1 pq where the maximum input signal dependence on the num- 53
54 (10) ber of scalar compandor quantization levels N was ascer- 54
55 Assuminig Laplacian distribution of the input speech sig- tained for the input speech signal modeled by the Lapla- 55
56 nals [7] cian distribution 56
57

2 |x|2 3 57
58 p(x, ) = e , (11) xmax = ln(N 2) , (14) 58
2 2
 Journal of ELECTRICAL ENGINEERING 62, NO. 1, 2011 5

01 01
02 02
03 03
04 04
05 05
06 06
07 07
08 08
09 09
10 10
11 11
12 12
13 13
14 14
15 Fig. 8. Theoretical results: Signal to quantization noise ratio as a Fig. 9. Comparison of the theoretical results with the G.712 stan- 15
16 function of the frame variance dard 16
17 17
18 18
one can complete the design procedure, described in the of implementation of the log-uniform quantizer having
19 19
previous section. Ng = 32 quantization levels in the proposed algorithm,
20 20
Let us now define the bit rate R corresponding to the one can notice two prominent picks corresponding to the
21 21
proposed coding algorithm frame lengths M = 8 and M = 33 , respectively. Since
22 22
for the both frame lengths the same SQN R has been
23 RNg 23
R = RN + . (15) achieved (SQN Raex = 36.44 dB), we have assumed the
24 24
M frame length M = 33 as a better rate quality solu-
25 25
tion for coding of the considered speech signal. For the
26 Take a notice of the fact that by RN , we have denoted 26
27
assumed frame length M = 33 , we have ascertained the 27
the number of bits per sample required for quantizing by
28
SQN Rpex vicissitude through the frames (see Figure 5). 28
N -level PLOC. Further, by RNg , we have denoted the
29 Similarly, for the log-uniform quantizer having Ng = 64 29
number of bits per frame having length M that is re-
30 quantization levels we have noticed three prominent picks 30
quired for the quantizing by Ng -level log-uniform quan-
31 tizer. From the last equation one can deduce that the at the frame lengths M = 11 , M = 29 and M = 49 31
32 decrease of the frame length M results in an unwanted (see Figure 6), which correspond to the highest SQN R 32
33 increase of the bit rate R . (SQN Raex = 36.42 dB) that can be achieved in the con- 33
34 sidered case of the proposed algorithm. Again, assum- 34
35 ing the highest frame length in order to provide rate 35
36 4 NUMERICAL RESULTS quality compromise, we have provided the SQN Rpex vi- 36
37 cissitude through the frames having length M = 49 (see 37
What is presented and discussed in this section are Figure 7). Additionally, assuming the appropriate com-
38 38
the theoretical and the experimental results that we have promise frame lengths (M = 33 in case of Ng = 32 and
39 39
achieved with the proposed coding algorithm, as well M = 49 in case of Ng = 64 ) we have provided the theo-
40 40
as the benefits over the performance achievable by the retical results (see Figures 8 and 9). From these figures,
41 41
coding solution designed according to G.711 standard. one can conclude that the proposed algorithm at the con-
42 42
Namely, we have considered forward adaptive PLOC
43 sidered bit rates completely satisfies the G.712 standard, 43
quantizer having L = 8 segments in the first quadrant
44 since it provides overreaching the curve defined by the 44
and N = 128 quantization levels. Moreover, we have
45 G.712 standard in the whole of variance range. From the 45
considered the log-uniform quantizer for gain quantizing
46 experimental results (see Figures 5 an 7), which have been 46
having Ng = 32 and Ng = 64 quantization levels, re-
47 spectively. Note that we have assumed 40 dB range of obtained for the appropriate compromise frame lengths, 47
48 the frame variances [7] while, according to Eqs. (7) and one can ascertain that the proposed algorithm provides 48
49 (8), we were obtaining the theoretical results. In addi- the average SQN R which is about 3.5 dB greater than 49
50 tion, we have been disposed of W = 10200 real speech the average SQN R achievable by the G.711 coding so- 50
51 samples (8 kHz speech) while, according to Eqs. (9) and lution SQN RaG.711 = 32.94 dB [9]. Additionally, one can 51
52 (10), we were obtaining the experimental results. It is highlight the fact that, in comparison to the solution de- 52
53 important to point out that, as it has been demonstrated signed according to the G.711 standard [9], our algorithm 53
54 in [1, 7, 8, 17, 18], the choice of the suitable frame length provides not only higher level of the average signal to 54
55 M is not a simple task and it usually results from the quantization noise ratio, but also performs the reduction 55
56 rate quality compromise (or rate SQN R compro- of the PCM bit rate for about 1 bits/sample. Therefore, 56
57 mise). Regarding the dependance of the SQN Raex on the we can reasonable believe that our algorithm provides 57
58 frame length M (see Figure 4), obtained in the case imposing benefits over the G.711 solution. Finally, since 58
 6 J. Nikolic Z. Peric D. Antic A. Jovanovic D. Denic: LOW COMPLEX FORWARD ADAPTIVE LOSS COMPRESSION . . .

01 the complexity of the quantizer implemented in the pro- [5] VASS, J.ZHAO, Y.ZHUANG, X. : Adaptive Forward-Back- 01
02 posed algorithm is the least possible we believe that the ward Quantizer for Low Bit Rate High Quality Speech Coding, 02
IEEE Transactions on Speech and Audio Processing 5 No. 6
03 proposed low complex algorithm is suitable for practical 03
(Nov 1997), 552557.
04 use. [6] JOHNSON, M. H.ALWAN, A. : Speech Coding: Fundamen-
04
05 tals and Applications, Wiley Encyclopedia of Telecommunica- 05
06 tions 5 (J. Proakis, ed.), John Wiley & Sons, New York, 2002, 06
07 5 CONCLUSION pp. 23402359. 07
08 [7] GERSHO, A.GRAY, R. M. : Vector Quantization and Signal 08
Compression,, Kluwer Academic Publishers, Boston, Dordrecht,
09 This paper has demonstrated the significance of the 09
London, 1992, Chapter 5, pp. 133172.
10 novel coding algorithm development via the gains in the [8] JAYANT, N. S.NOLL, P. : Digital Coding Of Waveforms, 10
11 compression and the performance that have been ascer- Principles and Applications to Speech and Video, Prentice Hall 11
12 tained by applying the proposed algorithm in speech cod- Secondary Education Division, New Jersey, 1984, Chapters 4-5, 12
13 ing over the G.711 standard [9]. Particularly, by com- pp. 115251. 13
14 paring the theoretical results obtained for the assumed [9] ITU-T, Recommendation G.711, Pulse Code Modulation (PCM) 14
of Voice Frequencies, International Telecommunication Union,
15 Laplacian source having a wide variance range with the 1972.
15
16 SQN R characteristic defined by the G.711 standard in [10] ATAL, B. S. : The History of Linear Prediction, IEEE Signal 16
17 the same variance range, it has been revealed that the Processing Magazine 23 No. 2 (March 2006), 154161. 17
18 proposed algorithm, along with about 1 bits/sample com- [11] RAMALHO, M. : Ramalho G.711 Lossless (RGL) Codec White- 18
19 pression, provides a higher level of SQN R which can be paper, Cisco Systems, Inc., 2002. 19
20 considered via the gain in the average SQN R of about [12] ORTEGA, A.VETTERLI, M. : Adaptive Scalar Quantization 20
1.62 dB. As with the theoretical results, the experimental without Side Information, IEEE Transactions on Image Process-
21 21
ing 6 No. 5 (May 1997), 665676.
22 results obtained by processing the real speech signal have 22
[13] NIKOLIC, J.PERIC, Z.POKRAJAC, D. : Average Com-
23 revealed the average SQN R gain of about 3.5 dB over plexity Analysis of Scalar Quantizer Design, Proc. of the 6th 23
24 the G.711 standard. Since it has been ascertained that WSEAS International Conference on Telecommunications and 24
25 the proposed coding algorithm satisfies the G.712 stan- Informatics, TELE-INFO 07, Dallas, Texas, USA, 2224 March, 25
dard [14] in the whole of the considered variance range, 2007, pp. 2227.
26 26
[14] ITU-T, Recommendation G.712, Transmission Performance
27 one can believe that the proposed algorithm will find its 27
Characteristics of Pulse Code Modulation (PCM), International
28 practical implementation in the high quality coding of Telecommunication Union, 1992. 28
29 signals, represented with less than 8 bits/sample, which [15] NIKOLIC, J.PERIC, Z. : Lloyd-Maxs Algorithm Implemen- 29
30 as well as speech signals follow Laplacian distribution and tation in Speech Coding Algorithm based on Forward Adaptive 30
31 have the time varying variances [17, 18]. Technique, Informatica 19 (2008). 31
32 [16] PERIC, Z.NIKOLIC, J. : Analysis of Compressor Functions 32
for Laplacian Sources Scalar Compandor Construction, Data
33 Recording, Storage and Processing 8 No. 2 (June 2006), 1524.
33
34 References 34
[17] MINOLI, D. : Voice over MPLS Planning and Designing Net-
35 works, McGraw-Hill., 2002, Chapters 1-2, pp. 1134. 35
[1] CHU, W. C. : Speech Coding Algorithms, Foundation and Evo-
36 [18] HERSENT, O.PETIT, J.GURLE, D. : Beyond VoIP Pro- 36
lution of Standardized Coders, John Wiley & Sons, New Jersey,
37 tocols Understanding Voice Technology and Networking Tech- 37
2003, Chapters 5-6, pp. 143183.
38 niques for IP Telephony, John Wiley & Sons, New Jersey, 2005, 38
[2] MOFFAT, A.TURPIN, A. : Compression and Coding Algo- Chapters 1-2, pp. 188.
39 rithms, Kluwer Academic Publishers, Norwell, MA, USA, 2002. 39
40 [3] RABINER, L. R.SCHAFER, R. W. : Introduction to Digital Received 9 June 2009 40
41 Speech Processing, Foundations and Trends in Signal Process- 41
42 ing, vol. 1, Jan 2007, pp. 1194. Jelena Nikolic 42
43 [4] HUANG, W.HO, C. C. : Nonstationary Linear Prediction Zoran Peric 43
Analysis of Speech Codec Corrected by Prestage Forward Vol-
44 Dragan Antic 44
ume Normalizer, Proc. International Symposium on Communi-
45 cations and Information Technologies, ISCIT 07 Sydney, Aus- Aleksandra Jovanovic 45
46 tralia, 1719 Oct 2007, pp. 15561560. Dragan Denic 46
47 47
48 48
49 49
50 50
51 51
52 52
53 53
54 54
55 55
56 56
57 57
58 58