1、 - - STD-ITU-T RECMN 6.727 APPENDIX I-ENGL 1994 m 48b2591 Ob25442 882 m - INTERNATIONAL TELECOMMUNICATION UNION ITU-T TELECOMMUNICATION STANDARDIZATION SECTOR OF TU Appendix 111 (Rec. G.726) Appendix II (Rec. G.727) (05/94) GENERAL ASPECTS OF DIGITAL TRANSMISSION SYSTEMS COMPARISON - OF ADPCM ALGORI
2、THMS Appendix 111 to ITU-T Recommendation G.726 Appendix II to ITU-T Recommendation G.727 (Previously “CCITT Recommendation”) FOREWORD The ITU-T (Telecommunication Standardization Sector) is a permanent organ of the International Telecommunication Union (IV. The -T is responsible for studying techni
3、cal, operating and tariff questions and issuing Recommen- dations on them with a view to standardizing teIecommunications on a worldwide basis. The World Telecommunication Standardization Conference (WTSC), which meets every four years, establishes the topics for study by the lTUT Study Groups which
4、, in their turn, produce Recornmendations on these topics. The approval of Recommendations by the Members of the ITU-T is covered by the procedure laid down in WTSC Resolution No. 1 (Helsinki, March 1-12, 1993). Appendix III to -T Recommendation G.726 and Appendix II to Recommendation ITU-T G.727 we
5、re prepared by -T Study Group 15 (1993-1996) and were approved on the 16 of May 1994. NOTE In this Recommendation, the expression “Administration” is used for conciseness to indicate both a telecommunication administration and a recognized operating agency. - O 1995 AI1 rights reserved. No part of t
6、his publication may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying and microfilm, without permission in writing from the . - - . STD*ITU-T RECMN 6.727 APPENDIX I-ENGL 1994 4Bb257L Ob25444 b55 . f. CONTENTS Background Overview of APCM algorithm
7、s Principles of Recommendations G.726 and G.727 and COM XVm-102 . 3.1 Adaptive Prediction and Reconstruction of the Sign al . 3.2 Adaptive Quantizer . 3.3 Quantizer Scale Factor Adaptation . 3.4 Adaptation Speed Control . finciples of COM XVIII-101 . 4.1 Prediction 4.2 The Futed AR Filter 4.3 Adapti
8、ve AR Filter . 4.5 Quantizer Adaptation ADPCM Decoder . 5.2 Synchronous Coding Adjusunent . Objective EvaIuation of ADPCM 6.2 Pedorrnance of the ADPCM Algorithms for Voiceband Data . Subjective Evaluation of ADPCM . 7.1 Subjective Evaluation of 32 kbit/s ADPCM . 7.2 Subjective Evaluation of G.721 Ex
9、tensions . 7.3 Subjective Evaluation of Embedded ADPCM 4.4 Update Equations 5.1 General Description 6.1 Theoretical Background 6.3 Objective Measurements Appendix I . References . Page 1 1 10 10 13 16 15 20 20 23 24 25 29 31 31 31 32 32 33 34 37 37 37 37 37 39 . AppSWG.726 . App . IVG.727 (05794) i
10、STD-ITU-T RECMN 6.727 APPENDIX I-ENGL Ob25445 591 m App. IIVG.726 - App. WG.727 Appendix II to ITU-T Recommendation G.726 and Appendix II to ITU-T Recommendation G.727 COMPARISON OF ADPChlI ALGORITHMS (Cerieva, 1994) 1 Background During the period 1982-1990, the CCIT adopted several adaptive differe
11、ntial pulse code modulation (ADPCM) algorithms. First, the 32 kbitls (ADPCM) algorithm described in Recommendation G.721 126; 61 was approved. Later on, Recommendation G.721 was extended with Recommendation G.723 to 40 kbits to support voice band data modems at the rate of 9.6 kbids, and to 24 kbids
12、 to allow reduction of the bit rate in cases of network conjestion 27. Prior to the definition of Recommendation G.723, other ADPCM algorithms of performance similar to the 40 kbit/s algorithm had been incorporated in DCME designs and used in telecommunications networks. These al_oorithms, which may
13、 be considered by bilateral agreement, are described in COM Xwr-i01 and COM XViIi-102 of the 1984-1988 study period.1) Finally, in July 1990, the CCfIT combined RecommendationsG.721 and G.723 and added operation at 16 kbit/s for overload situations. The combination resulted in a new Recommendation G
14、.726. The CCITT also approved the embedded ADPCM algorithms of RecommendationG.727, which are extensions of the fixed rate ADPCM algorithms defined in Recommendation G.726. This appendix presents a unified introduction to all these algorithms, their main features and their performance. Clause 2 give
15、s an overview of all ADPCM algorithms that the CCm has considered. Clause 3 reviews the principles of the algorithms of Recommendations G.726 and G.727 and COM Xvm-102. The principles of the aigorithm of COM XViIi-101 are described in clause 4. The remaining clauses outline the main subjective and o
16、bjective results for the performance of the various algorithms. 2 Overview of ADPCM aIgorithms Figures 1 and 2 show a simplified block diagram of a G.726 encoder and decoder, respectively. Figures 3 and 4 show a simplified block diagram of a G.727 encoder and a decoder, respectively. In each set, th
17、e coder consists of a logarithmic- to-linear PCM converter, an adaptive quantizer, an inverse adaptive quantizer, and an adaptive predictor. The PCM converter converts the A-law or p-law PCM input signal s(k) to a uniform PM signai (k is the sampling index for a sampling period of 125 ps). The predi
18、cted estimate of the input signa1 s,(k) is subtracted from the uniform PCM signal, sl(k), to yield a difference signal d(k): 4k) = Si(k)-S,(k) (1) The difference signal is then transformed into a logarithmic presentation with the base 2 and scaled by a scale factor y(k) In Recommendation G.726, the
19、quantizer used is a 31-, 15, 7- or 4-level non-uniform adaptive quantizer that stops adapting in the presence of a stationary input. This enhances the performance for voiceband data signals. In Recommen- dation G.727, the adaptive quantizer has 32, 16, 8 or 4 levels. Either quantizer codes the signa
20、l d(k) into I(k), a code word of 5.4.3 or 2 bits respectively, with one bit always for the sign. that is computed as described below. L ) The US patent that describes the final algorithm has some differences from the algorithm described in COM XVIII-102 53. ,App. IIuG.726 - App. IYG.727 (0994) 1 - S
21、TD-ITU-T RECMN 6.727 APPENDIX I-ENGL 1999 m 48b259L Ob259qb q28 m -. i t App. IWG.726 - App. WG.727 (0-4) STD-ITU-T RECMN G-727 APPENDIX I-ENGL 1994 48b2571 Ob25447 3b4 - - 4 App. IIuG.726 - App. WG.727 (05194) 3 STD-ITU-T RECMN 6.727 APPENDIX I-ENGL 1994 9 48b259L Ob25448 ZTCJ - -. 4 App. mVG.726 -
22、 App. WG.727 (05/94) t- + c The main difference between the fixed ADPCM algorithms of Recommendation G.726, COM XVIII-IO1 and COM XVIII-103 on the one side. and the embedded algorithms of Recommendation G.727 on the other side is as follows. In the embedded algorithms of Recommendation G.727,- the d
23、ifference between the input and the estimated signal is quantized into code words consisting of enhancement bits and core bits. The core bits are used for prediction, both in the encoder and the decoder, while the enhancement bits are used to reduce the quantization noise in the reconstructed signal
24、. Thus, the core bits must reach the decoder to avoid mistracking, but the enhancement bits can be discarded, if such bit-dropping can alleviate congestion. The embedded ADPCM algorithms can operate with 5-, 4-, 3- and 2-bits per sample in their feedback path (Le. at rates of 40, 32, 24 and 16 kbids
25、) with conversion to and from 6.F kbids A-law or p-law PCM channels. These G.727 algorithms are referenced as (x,y) pairswhere x refers to the feed-forward ADPCM bits and y refers to the feedback ADPCM bits. For example, (5J) represents 40 kbit/s embedded algorithm with two core bits, Le. with a min
26、imum bit rate is 16 kbit/s. The impetus for developing Recommendation G.727 was to provide a flexible way to alleviate congestion at any point in a packet network without the need for exchanging control messages between the various nodes in the backward path of the connection (i.e. towards the trans
27、mitter). This avoids the “freeze-out” associated with fixed rate ADPCM coding, when transmission capacity is not available and the leading edge of speech bursts are clipped (Le. the beginnings of words are chopped). This is important when the end-to-end path includes multiple nodes. The difference b
28、etween the various fixed rate ADPCM algorithms resides in the way that they accommodate 9.6 kbit/s modem signals. In Recommendation G.726,40 kbitfs ADPCM is used for voiceband data, while 32 kbit/s ADPCM is used for speech. Accordingly, a bypass arrangement is needed so that upon detection of voiceb
29、and data, the appropriate coding is applied without affecting the coding for speech 9. COM XWI-101 uses the same 32 kbit/s ADPCM algorithm for speech as well as voiceband data 37. The predictive strucrure, which is different from that of all other ADPCM aigorithms, is composed of a 10th order adapti
30、ve zero predictor, a 4th order adaptive-pole predictor, a 16th-order fixed-pole predictor and an offset predictor. The adaptive pole filter is reserved for highly correlated signals such as speech, while the fixed pole filter is for voiceband data. The relative contribution of each filter is regulat
31、ed by a set of adaptive gain coefficients. By controiling three different filters within the same structure, the algorithm treats speech signals and voiceband data modem signals up to 9.6 kbit/s with 32 kbids ADPCM; it does not require a change-over between speech and voiceband data at 9.6 kbitfs; t
32、he price is additional complexity. The adaptive quantizer of this algorithm operates in a 4-bit quantization mode and does not use a tone and transition detector. Figures 5 and 6 give the block diagrams of the encoder and decoder of the COM XWI-101 algorithm. COM XVm-102 uses a special 32 kbit/s BPG
33、M algorithm that uses 5 bits/sample and is optimized for voiceband data 33; 531. Following detection of a 2100 Hz tone, the linear PCM bit stream is down-sampled from 8 kHz to 6.3 kHz through a 100-tap symmetric finite impulse response interpolating filter. This interpolating filter introduces a fla
34、t delay of 6 ms equally distributed between the encoder and the decoder. To maintain the overall line rate of 32 kbit/s, the ADPCM coding uses 5 bits. To avoid aliasing, the inputs bandwidth must be limited to 3.2 kHz. Also, a realignment from a 6.4 kHz x 5 structure to an 8 kHz x 4 structure is req
35、uired. The corresponding encoder and decoder block diagrams are shown in Figures 7 and 8, respectively. In these figures, the tone detector block is assumed, because it is not described in the available documents from the algorithm developers 33; 531. . The adaptive predictor relies on the whole cod
36、eword I(k) for Recommendation G.726 and the fixed rate ADPCM algorithms, and on the core codeword I therefore, it does not exhibit the synchronous tandem property described in 5.2. Clause 5 recapitulates the encoder principles and explains the differences among the various algorithms of Recommendati
37、ons G.726 and G.727 and COM Xvm-102. Discussion of the algorithm of COM XVLII-101, whose stnicture is different from the other algorithms, is the subject of clause 6. 3 Principles of Recommendations 6.726 and G.727 and COM XVIII-102 3.1 The primary function of the adaptive predictor is to compute th
38、e signai estimate s b: i = 1, ., 6 the moving average coefficients for sample k; are the autoregressive coefficients for sample k; is the diference signal ar sample k; is the quantized difference signal at sample k; is the quantization error at sample k; is the normalized quantizer output for the in
39、put x; = I x I -y) in the logarithmic domain; is the the scale factor error at sample k dq (k) e (4 Qixl Y(k) 10 App. IIYG.726 - App. WG.727 (05194) STD*ITU-T RECMN 6.727 APPENDIX I-ENGL 1779 98b257L Ob25955 430 I -. The startins values are: d(0) = s,(O) = s,(O) = O and dq(k) = O for k (11) for i =
40、1,2, ., 6. Note that i bi(k) I 5 2. As will be seen later, when a transition is detected the predictor coefficients are reset, i.e. o:= $= O, and b/= O for i = 1,2, ., 6. As above, sgn O = 1, and sgn d,(k) = O, fork e O For 40 kbit/s coding, the adaptive predictor is changed to decrease the leak fac
41、tor used for zeroes coefficient operation. In this case, equation (1 1) becomes: Note that bi(k) is implicitly limited to k 2. 3.1.2 COM Xvm-102 In the 32 kbit/s algorithm of COM XVIII-102 33, the coefficient update equations are of the same form but the leak factor is lower for both the poles and t
42、he zeros. ?Tie corresponding equations are: = (1 - 2-19 + 2-9 sgn P()I SP p(k - 111 o (7 3 and bk = (1 - 2-*O) bf-l + 2-9 sgn dq(k) sgn d,(k - i) 12 App. IwG.726 - App. WG.727 (05/94) (9) 3.2 Adaptive Quantizer Normalized Quantizer Input Range I I(k) I log2 I 44 I - (Q 4.31,+- ) 15 4.12, 4.31) 14 3.
43、91, 4.12) 13 3.70, 3.91) 12 3.47, 3.70) 11 3.22, 3.47) 10 2.95, 3.22) 9 2.64. 2.95) 8 2.32, 2.64) 7 1.95, 2.32) 6 1.54, 1.95) 5 1.08, 1.54) 4 0.52, 1.08) 3 4,13. 0.52) 2 4.96, 4.13) 1 (-w, 4.96) O All ADPCM algorithms have non-uniform midrise adaptive quantizers that are based on the minimum-mean-sq
44、uared error Lloyd-Max quantizer at 32 kbirfs 35; pp. 131-1341. They operate in a bimodal fashion (slow and fast) with an adaptive scale factor. y(k), to accommodate both speech and voice-band data signals 47. In Recommendation G.726. the quantizer is a 31-, 15, 7- or 4-level non-uniform adaptive qua
45、ntizer for operation at 40, 32, 24 or 16 kbirfs, respectively. Each rate has it own separate quantizer. In Recommendation G.727. a 32-, 16-, 8- or 4-level non-uniform adaptive quantizer used to quantize the difference signal. d(k). The various quantizer tables are embedded within each other so that
46、the decision levels are forcibly aligned to ensure that the decision levels for the 32, 24 and 16 kbit/s quantizers are subsets of those for the 40 kbit/s quantizer. This contrasts with the algorithms of Recommendation G.726 where the decision leveis are not aligned which makes them unsuitable for e
47、mbedded applications, but with a slightly improved signai to quantization noise ratio as shown 8.3.1. Tables I, 2 and 3 give the inpuiloutput normalized characteristics of the G.726 quantizer for operation at 40, 32 and 24 kbit/s. Table 4 corresponds to the algorithm of COM Xvm-102 33. Normalized Qu
48、antizer output log2 I dq(k) I - Y(k) 4.42 4.21 4.02 3.81 3.59 3.35 3.09 2.80 2.48 2.14 1.75 1.32 0.8 1 0.22 4.52 -m TABLE 1 Quantizer Normalized Inpu/Output Characteristic for 40 kbis Operation of Recommendation G.726 Tables 5 and 6 give respectively the normalized input and output Characteristic (i
49、nfinite precision values) of the quantizer for the G.727 algorithms with 5,4,3 and 2 core bits. Both tables show the alignment of the decision levels for the various bit rates. In all these tables, the most significant bit is the sign bit and the remaining bits represent the masnitude. The 5, 4-, 3- or 2-bit quantizer output, Z(k), of Table 6 forms the 40, 32, 24 or I6 kbiils output signal that comprises both the enhancement and core bits. The 16 kbit/s algorithm is
