1、 STD*ITU-T RECMN G-72b APPENDIX I-ENGL L97Li 48bZ.571 0b258bb 7LiT INTERNATIONAL TELECOMMUNICATION UNION ITU-T TELECOMMUNICATION STANDARDIZATION SECTOR OF ITU Appendix 111 (Rec. G.726) Appendix II (Rec. G.727) (05/94) GENERAL ASPECTS OF DIGITAL TRANSMISSION SYSTEMS COMPARISON - OF ADPCM ALGORITHMS A
2、ppendix 111 to ITU-T Recommendation G.726 Appendix II to ITU-T Recornmendation G.727 (Previously ?CCITT Recommendation?) FOREWORD The -T (Telecommunication Standardization Sector) is a permanent organ of the International Telecommunication Union (ITU). The ITU-T is responsible for studying technical
3、, operating and tariff questions and issuing Recommen- dations on them with a view to standardizing telecommunications on a worldwide basis. The World Telecommunication Standardization Conference (WTSC), which meets every four years, establishes the topics for study by the ITU-T Study Groups which,
4、in their turn, produce Recommendations on these topics. The approval of Recommendations by the Members of the -T is covered by the procedure laid down in WTSC Resolution No. 1 (Helsinki, March 1-12, 1993). Appendix Ei to T-T Recommendation G.726 and Appendix II to Recommendation -T G.727 were prepar
5、ed by ITU-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. c o m 1995 All rights reserved. No part of this
6、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 ITU. CONTENTS Background Overview of ADPCM algorithms Principles of Recommendations G.726 and G.727 and COM XVI-102 . 3.1 Adap
7、tive Prediction and Reconstruction of the Sign al . 3.2 Adaptive Quantizer. 3.3 Quantizer Scale Factor Adaptation . 3.4 Adaptation Speed Control . Principles of COM XVIII-101 . 4.1 Prediction 4.2 The Fixed AR Filter 4.3 Adaptive AR Filter . 4.4 Update Equations 4.5 Quantizer Adaptation ADPCM Decoder
8、 . 5.2 Synchronous Coding Adjustment . 5.1 General Description Objective Evaluation of ADPCM 6.1 Theoretical Background 6.3 Objective Measurements 6.2 Performance of the ADPCM Algorithms for Voiceband Data . Subjective Evaluation of ADPCM . 7.1 Subjective Evaluation of 32 kbitls ADPCM . 7.2 Subjecti
9、ve Evaluation of G.721 Extensions . 7.3 Subjective Evaluation of Embedded ADPCM Appendix I . References . Page 1 1 10 10 13 16 18 20 20 23 24 25 29 31 31 31 32 32 33 34 37 37 37 37 37 39 . App . IIUG.726 . App . IVG.727 (05194) i - _ STD-ITU-T RECMN G.72b APPENDIX I-ENGL I774 9Bb2591 Ob25Bb9 b57 App
10、. IIUG.726 - App. IyG.727 Appendix II to ITU-T Recommendation G.726 and Appendix II to ITU-T Recommendation G.727 COMPARISON OF ADPCM ALGORITHMS (Geneva, 1994) 1 Background During the period 1982-1990, the CCIT adopted several adaptive differential pulse code modulation (ADPCM) algorithms. First, th
11、e 32 kbits (ADPCM) algorithm described in Recommendation G.721 26; 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 kbit/s. and to 24 kbids to allow reduction of the bit rate in cases of network co
12、ngestion 27. Prior to the definition of Recommendation G.723, other ADPCM algorithms of performance similar to the 40 kbids algorithm had been incorporated in DCME designs and used in telecommunications networks. These algorithms, which may be considered by bilateral agreement, are described in COM
13、xvDI-101 and COM WEI-102 of the 1984-1988 study period.) Finally, in July 1990, the CCITT combined RecommendationsG.721 and G.723 and added operation at 16 kbitls for overload situations. The combination resulted in a new Recommendation G.726. The CCITT also approved the embedded ADPCM algorithms of
14、 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 gives an overview of all ADPCM algorithms that the CCIT has consid
15、ered. Clause 3 reviews the principles of the algorithms of Recommendations G.726 and G.727 and COM Xvm-102. The principles of the algorithm of COM XVIIi-101 are described in clause 4. The remaining clauses outline the main subjective and objective results for the performance of the various algorithm
16、s. 2 Overview of ADPCM algorithms 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, the coder consists of a logarithmic- to-linear PCM converter, a
17、n 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 PCM signal (k is the sampling index for a sampling period of 125 ps). The predicted estimate of the input signal s,(k) is subtracted from t
18、he uniform PCM signal, sl(k), to yield a difference signal d(k): 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 quantizer used is a 31-, 1.5, 7- or 4-level non-uniform adaptive quantizer that st
19、ops 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 signal d(k) into I it does not require a change-over between speech and voiceband data
20、 at 9.6 kbitfs; the 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-IO1 algorithm. COM XVm-102 uses a spec
21、ial 32 kbit/s mPCh4 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 6.4 kHz through a 100-tap symmetric finite impulse response interpolating filter. This interpolating filte
22、r introduces a flat 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 kIiz. Also, a realignment from a 6.4 kHz x 5 structure to an 8 kIiz
23、x 4 structure is required. 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 relie
24、s on the whole codeword 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 algorith
25、ms of Recommendations G.726 and G.727 and COM XVIII-102. Discussion of the algorithm of COM XVIII-101, whose structure 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 Adaptive Prediction and Reconstruction of t
26、he Signal The primary function of the adaptive predictor is to compute the signal estimate se) from the quantized difference signal dq(k) and past values of the reconstructed signai se(k). The predictor has the form of an autoregressive moving average (ARMA) filter whose frequency spectrum fits a wi
27、de range of input voice-band signals. The signal estimate is computed from the reconstructed signal s,(k) as: with and where are the autoregressive coefficients for sample k; the moving average coefficients for sample k; is the difierence signal ar sample k; is the quantized difference signal at sam
28、ple k; is the quantization error at sample k; is the normalized quantizer output for the input x; = I x I - y(k) in the logarithmic domain; is the the scale factor error at sample k 10 App. IIYG.726 - App. IUG.727 (0994) The starting values are: d(0) = s,(O) = s,(O) = O and d 81. This slow adaptatio
29、n causes some instabilities for the Bell 202 modems and the CClT V.23 series of modems, where the mark and space tones are wide apart, when they operate in the character mode. Modifications were introduced to force the quantizer in the fast adaptation mode, and to reset the predictor coefficient, wh
30、en an FSK signal is present. In addition, 15-level quantization is adopted instead of the original 16-level quantization to allow the use in US Networks that do not provide bit sequence independence. Thus, the predictor coefficients are reset 18 App. IIYG.726 - App. IIlG.727 (05/94) - STD-ITU-T RECM
31、N G.72b APPENDIX I-ENGL L99q qBb259L Ob25887 b7Y to O and the quantizer is unlocked, i.e. (up = l), following a transition between two tones 7; 341. A tone is detected if (k) 24 rdk) = O, otherwise. To summarize, the intermediate variable up) is defined as follows: ap(k - 1) , otherwise, and up) = u
32、p fork I O Thus, up(k) i, 2 whenever one of the following conditions is true: dmc.(k) - dmk) is large because the average magnitude of I 31; 361 and on personal correspondence with Mr. Atsushi Shimbo from OKI Electric. 4.1 Prediction The transfer function H(B) of the composite adaptive predictor of
33、COM XWI-101 is given by: where (E) is the transfer function of the 10th-order adaptive moving average, a 4th degree adaptive autoregressive (AR) predictor; a 16th degree fixed autoregressive (AR) predictor; and As explained earlier, the adaptive AR(4) filter is for speech signals while the fixed AR(
34、16) filter is for voice band signals. The order of the fixed predictor is so chosen because the prediction gain tends to saturate above the 16th order 31. The adaptive gain gap is chosen such that it tends to 1 for speech signals and to O for voiceband data. Inversely, the fixed gain gfi is chosen s
35、uch that it tends to O for speech signals and to 1 for voiceband data. Thus, the signal estimate se, is computed from four components q(k) as follows: In this equation, el(k) is the output of the MA(10) predictor whose coefficients for the kth sample are b: i = (I, ., lo), q(k) is the output of the
36、adaptive AR(4) prsdictor with the coefficients at k q; i = (1, ., 4), q(k) is the output of the fixed AR(10) filter whose fixed coefficients are denoted as ci, i = (1, ., 16), and the offset ed(k) is a filtered version of the quantized difference d Ziand zf;: - 2f-l Zi + 1 (43) (1 - 2-7)2;-1 - zi +
37、zi, otherwise, 2; = 26 App. IWG.726 - App. WG.727 (05194) STD=ITU-T RECMN Gm72h APPENDIX I-ENGL 1994 m 4862593 0625875 740 m Root Location Variable Value where Frequency (Hz) Value 26 = 0,z: = 7t) The initial values and the corresponding frequencies are given in the following tables: I RootLocation
38、I Frequency(Hz) I Value 0.3501 0.3501 0.5520 1.5010 2.2160 3.1416 (x) Variable Value O 445.8 702.8 1911.1 2821.5 4000 O. 1570 0.0313 0.03 13 0.03 13 40 Thus, the conditions for updating the location of the roots can be expressed as shown in the following table: - App. IWG.726 - App. IyG.727 (05194) 27
