CONTENTS - Introduction - University of Hertfordshire
Contents
1. e Kondoz M A 1994 Digital Speech Coding for Low Bit Rate Communication Systems Surrey John Wiley amp Sons e Owens J F 1993 Signal Processing of Speech London The Macmillan Press LTD e Pree H William Teukolsky A Saul Vetterling T William Flannery P Brian 1992 Numerical Recipes in C The art of scientific computing Second Edition Cambridge Cambridge University Press e Analog Devices 1999 Getting Started guide Norwood USA Analog Devices Inc Computer Products Division Onetechnology Way e Analog Devices 1999 ADSP 21065L EZ LAB Development System Manual page 1 6 Shrivenham technology International Europe Limited e Analog Devices 1998 ADSP 21065L SHARC Technical Reference Norwood USA Analog Devices Inc Computer Products Division One technology Way e Analog Devices 1998 ADSP 21065L SHARC User s Manual Norwood USA Analog Devices Anc Computer Products Division One technology Way MSc Project Report 39 Echo cancellation using ADSP 21065L APPENDIX Result graphs The result graphs where obtained from a digital oscilloscope with persistence of 1 second The grey areas show the various positions the signals took within a 1 second time period The bold black lines depict the signals positions at the moment of print request The part of the graph with the 1 at the left side denotes channel 1 of the oscilloscope while with the 2 at the left side denotes channel 2 of the oscil2. h E SA ART SR E E TA E A O TA D H s H H H H H Mean a 42 4mY CH1 1 00Y CH 1004 M 250m5 CHI 240mY Figure 19 Channel 1 displays the noisy system input d n while channel 2 displays the system output e n Ideally this should be a pure sine wave MSc Project Report 44 Echo cancellation using ADSP 21065L The following two graphs display on channel 1 of the oscilloscope the ideal sinusoidal output of the system while on channel 2 the actual system output It must be noted that the first graph printout was generated while the coefficients start to adapt Pom their initial values zero while the second graph printout was generated after allowing some time for the coefficients to adapt Tek P y Trig d M Pos 1 220ms MEASURE YAOI AAA INICIO IIS NOAA WK ENKER NOIA A s s Ka KEN AE A um AN NIK e A CN PA Kai KC D z Freg A 1 504kHz SA seas E CR oS EH 8 Mean 1 53mY PEEEEREEREETEEEEE EDEL EPREP ER EEEEEEEEEEEE ER EERE ORG RNG KK GEES A AA A CIN Mean 64 3mY CHT 1004 CH2 100 M250pus CHI 62 0m Figure 20 Channel 1 displays the ideal system output and channel 2 the actual system output just at the beginning of the adaptation procedure MSc Project Report 45 Echo cancellation using ADSP 21065L Tek JL Trig d M Pos 1 220ms MEASURE 1 520kH2 Freq 1 504kH2 Mean bh dm CH2 10
3. 1 f4 12 pm is m8 0 STEPSIZE e n tu n 12 w0 n lentr TAPS 1 do update _veights until lce 6 0 12 4 dm 10 m0 12 pm i8 m8 8 wi n l 4 u n 1 1 12 wi l n king working directories eking project dependencies and exporting Makefile Project files Status information Code editor Figure 6 This is a captured image of VisualDSP environment instance just after the talk through program was opened The above picture shows an instance ofthe Visual DSP environment By building the project VDSP checks all dependencies for the files included in the project and invokes the compiler which checks the code for improper declarations invalid instructions and syntax errors if all go well the linker is invoked next The linker then links all the files together to produce a dxe executable file This file is loaded on to the DSP processor by the Debugger environment An instance of which can be seen at the following image Fig7 MSc Project Report 16 Echo cancellation using ADSP 21065L Analog Devices VisualDSP Debugger Target ADSP 21065L EZ KIT Lite EZ KIT 21065L File Session View Register Memory Debug Settings Tools Window Demo Help screens aes os DataDM Untitled 00C000 00C001 00C002 00C003 00C004 00C005 00C006 00C007 00c008 00C009 00C00A 00C00B 00co00c 00C00D 00CODE 00C00F 00C010
4. 4 which is the mean square error where E is the expectation operation By successive substitution of equation 3 to 1 and 1 to 4 we yield the following equation c Ela n wr n R w n 2 w n p equation 5 Where R E Leit n is the auto correlation matrix and p E ld n x n which is the cross correlation vector By solving the following equation Ref which minimizes the mean square error we can acquire the desired values for the filter coefficients 2 p 2RW 0 5 w n as gt RW p W wo Wi W2 Wn 1 The above lead to the Wiener Hopf equation w R P equation 6 that gives the optimum weights for the filter Obtaining the filter weights thisway is time consuming and unrealistic due to the calculation of the auto correlation matrix and the cross correlation vector This is why he practical LMS gt used_in implementations adapts the weights on a sample by sample basis The formula involved is w n 1 w n u V n equation 7 w n is the weight vector u the adaptation step size and V n the true gradient vectors at the n sample u the adaptation step size also controls the rate of convergence and reliability of the system By increasing the value of u the algorithm MSc Project Report 22 Echo cancellation using ADSP 21065L updates the filter weights by larger steps which improves the system s response time but at the same time makes the weight values no
5. Referencial drid 38 10 0 Bibl st EE gt E daa pa 39 ARRENDAR e M i anneni 40 RES graphs seirene m aan ae aae E EE N eelere 40 APPENDIX Tia a en E 49 Program E tt RE Tah 49 MSc Project Report 2 Echo cancellation using ADSP 21065L ACKNOWLEDGEMENTS I wish to thank sincerely my project supervisor Dr Mazim Mahmoud for his continual help and encouragement throughout the project I would also like to thank Professor T Alukaidey and G Pissanidis for their invaluable help in understanding the inner workings of the DSP used in my project MSc Project Report 3 Echo cancellation using ADSP 21065L ABSTRACT This report presents an attempt to create an echo canceller using a DSP Digital Signal Processor chip The report goes through the steps taken to design the system and implement it on Analog Devices ADSP Y 21065L EZ KIT Lite The filter used in the implementation was a transversal FIR filter and the algorithm used for updating the adaptive filter coefficients was LMS Least Mean Square The system performs well for echo delays of up to 10ms with an attenuation of approximately 20dB of echo attenuation The number of filter taps used was 64 MSc Project Report 4 Echo cancellation using ADSP 21065L 1 0 INTRODUCTION 1 1 Chapter Overview This chapter explains the project giving information about its aims and objectives It also gives some background information on the tools used and the layout of the project 1 2 Project
6. and ways to eliminate the problem caused by the echo effect It briefly describes the operation of an echo suppressor and an adaptive echo canceller 2 2 What le Echo Echo as mentioned in the introduction is the delayed or distorted signal reflection that returns to its source There are mainly two types of echo the acoustic echo and the electrical echo 2 2 1 Acoustic Echo When we have a microphone in close proximity to a loudspeaker what usually results is acoustic coupling between the tw This problem was first noticed when the first telephones that offered a loudspeaker option for hands free conversations The speaker s voice is coupled through the listener s microphone and so an echo is created The following figure fig 1 illustrates the procedure Near End Far End Ab pA Speaker lt lt Listener Acoustic Coupling xA Far End Near End Listener O B Speaker B xA Figure 1 Echo generation by acoustic coupling at the near end speaker MSc Project Report 7 Echo cancellation using ADSP 21065L 2 2 2 Electrical Echo Electrical echo is the echo generated whenever a transmission line carrying a signal is not properly terminated In a telephone network the main source of electrical echo is at the hybrid when connecting a two wire transmission line telephone to a four wtre transmission line to allow for full duplex communication The following figure fig 2 illustrates the concept Two wire Fou
7. cancellers within a communications channel to achive full duplex echo cancellation 5 4 Adaptive Filter s Algorithms There are many type of algorithms developed for use in adaptive digital filters LMS Least Mean Square RLS Recursive Least Square NLMS Normalized Least Mean Square and many more Although not the best algorithm to utilize in an echo canceller most commercial applications favourthe LMS This is mainly because it is easy to implement and exhibits stability in performance In this project the LMS algorithm was therefore implemented based on the same criteria Although the NLMS was considered as a second algorithm implementation unfortunately time restrictions did not allow for any implementations other than the LMS The LMS algorithm changes the filter coefficients based on the minimization of the square error of the system As mentioned before the system cancels the echo present in the desired signal d n by predicting the echo Y n and subtracting it from the actual system output e n If the echo prediction Y n is not perfect then the subtraction result ein also gives the system error an constitutes of the output signal and the residual echo int 5 4 1 Mathematical Modelling So we have The filter output u W n X n Zil n i equation 3 the system output MSc Project Report 21 Echo cancellation using ADSP 21065L e n d n Y n equation 1 And the algorithm output e E e n equation
8. introduction The very concept of communication involves the presence of at least one speaker and a listener that interchange roles A speakers ask is to transmit to the listener some information data This data propagate through a medium usually either air or a cable conductor During this propagation of data part of it are reflected back to the speaker thus an echo is created IC the delay between the transmission of the data signal and the echo signal is small and at the same time the amplitude of the echo signal is small compared to the initial transmitted data then the echo signal is not noticeable If the delay exceeds a few tens of milliseconds and its amplitude is not negligible compared to that of the initial transmitted signal the echo is noticeable and presents a communication problem In voice data communications is a nuisance and in data communications can cause data corruption or be falsely mistaken as valid data itself This report focuses on the elimination of echo in communications by the use of an adaptive filtering technique 1 3 Project Aims and Objectives The objective of this project was to create an echo cancelling module for use mainly to deal with acoustic echo which results from the reflection of sound waves and acoustic coupling between the microphone and loudspeaker as well as with electrical echo generated the two to four wire conversion hybrid transformer due to network impedan
9. over 100 Math DSP and C runtime library routines an assembler linker loader simulator and EPROM Splitter The software had a licence of 45 days after which you could only run it in the limited DEMO edition MSc Project Report 15 Echo cancellation using ADSP 21065L VisualDSP Integrated Development Environment Lms asm 7 E el sjej ala dall ea sel alula clear_bufs dm i0 m0 0 pm i8 m8 0 clear delay line e weights rts lms alg r5 dm Right_Channel In dm Right Charnel Out r5 kalo sima r3 DM Left_Channel_In noise 19 15 33 DM Right_Channel Out r9 Lms asm i Talkthruasm f r0 31 rl 31 0 float r3 by r0 fl float r9 by rl dm i0 m0 0 f4 pm 18 m8 store u n in delay line 4 w0 n f6 0 f4 fO dm 10 m0 4 pm 18 m8 8 u n wO n 0 u n 1 f4 vl n El2 f0 F4 fO dm 10 m0 4 pm 18 m8 12 u n 1 w1 n 0 u n 2 4 w2 n lentr TAPS 3 do macs until lce macs l2 f0 f4 f6 f8 f12 fO dm 10 m0 4 pm 18 m8 12 u n 1 wi n 8 sum of prod 0 u n i 1 4 wi l n 12 f0 f4 8 f8 12 12 u n 14 1 wN 1 n 13 84 12 13 y n Wi r0 31 dh r4 fix 13 by r0 am Right_Channel Our r4 f6 1 13 f6 ein r0 31 r5 fix f6 by r0 DM Left_Channel Out r5 fl f6 f7 f4 dm 10 m0 fl STEPSIZE e n f4 u n f0 f
10. 0 M250ps CHI 62 0m CHT 1 004 Figure 21 Channel 1 displays the ideal system output and channel 2 the actual system output after the adaptation procedure advanced for some time approximately 5 seconds MSc Project Report 46 Echo cancellation using ADSP 21065L The following two graphs display the ideal output signal subtracted from the actual system output thus giving us the residual noise Ideally if the system adapted perfectly this should be a flat line The difference between the graphs is in the time scale used 250us for the first 2 50ms for the second Tek Ef Trig d M Pos 1 220ms MATH ARS ai a E TI rel ee WO YL SSL Va EC IA GG Was Sa EA E YON ORL Sc ae aN KEE AE AS a IN he T ee TA SA A COCR EIER ARI OSA AD Oe OO EE 8 SR TS WSN LONER e E et E e E E e e C E S S TE e ar KC E E a OF eee el Sw Ksmwga ae KKK KKK OS A KES SRIKKSEKS SKS Window 1 UT RW a CL Me EE nH Gee er ee Ee EE CRC Scorn Shae came Nee oe Eee a Samat E e Ae A CH1 100 CH2100 M 250us CH1 Z 62 0mY Figure 22 Actu lsystemoutput minus ideal system output MSc Project Report 47 Echo cancellation using ADSP 21065L Trig d M Pos 1 250ms MATH Tek Jl AMIENS A 8 wn yess yen ERR RES oh O ES SS sg NSS 8 en KKK a 6 6G OTe ra KEIER SKI OO ESO 8 08 Oe b S D H Sov sts ss RM es she EAN RRC OW RL DR gt gt 2 V M 2 Indow s H H H H H M anni A A Eege ST EE
11. 0 00000000 00098628807 00284010 Mii 00000000 B11 00000000 L11 00000000 0041955411 00008FFF Mi2 00000000 B12 0000001 112 00000000 0058826646 000081BC Mi3 00000000 B13 00000000 L13 00000000 00336 4B465 00000000 Mi4 00000001 B14 00200000 114 00000000 00008004 MiS FFFFFFFF B15 00000000 115 00000000 0080105234 15961725 Halted pp pee e 9 Plot window Disassembly window Data Memory window Status Console window Figure 7 This image of the debugger was captured while the echo canceller program was running on the DSP chip This utility debugger allows the user to step by step analyse the code that runs on the DSP to find problems in coding logic The debugger has all the necessary functions such as step by step instruction execution watch points break points disassembly window and many more useful features to examine the program and make the necessary corrections Lo make it work properly If during the debugging procedure the user identifies some error and wants to correct them he has to go back to the editor make the changes in the code and re compile re link and re debug the program until it works as intended The following diagram shows the main procedures involved Fig 8 MSc Project Report Echo cancellation using ADSP 21065L 17 Figure 8 Basic steps that lead to a working application software MSc Project Report Write Edit Code Compile E Link Execute Debug Application Echo can
12. 00C011 00c012 00C013 0 0 0 0 0 ply 00814F call Blink_LEDs_Test wait_fdrever 0081B0 call wf it_flagl pooco14 0 0081B1 bit tgl ustatl 0x3 00C015 0081B2 dn 0xPf ustatl ooco16 0081B3 jump frait_forever 00C017 waitjflagl o0co18 0081B4 if aal An jump wait_flagl ml 00C019 0 00C01A O 0081B5 00C01B O 0081B6 aocoic 0 00C01D 0 0081B7 00C01H 0 0081B8 weights 0 0004661364 0 0032796552 0 0015320795 0 0022660492 0 0014818615 0 0 0 0 0 0 0 0024987245 0 12229151 0 0033353905 0 0002048372 0 0020395853 0 0014807887 0 Fear eve BAO 2 2 E 53188866 CURLCNTR 1 NAN EE LCNTR 1 4012985E 045 0026343407 i 0012799388 LADDR 1 NAN 0013140995 004094739 00093453273 0 14339706 0000D27C MO FEFFFFFF B0 0000D266 00000020 0023513492 00002000 M1 00000000 B1 00040002 00000000 0036642605 00000000 M2 00000000 B2 00000000 00000000 0011311031 01000008 M3 00000000 B3 00000500 00000000 0000D03E M4 00000002 B4 00800004 00000000 0027901367 0000D060 MS 00000000 BS 08800004 00000000 00000000 00100000 M6 00000001 M FFFFFFFF B6 B 01900000 00800000 00000000 00000000 00081204053 DAG2 PM oooocooo M8 00000001 B8 oooocooo M9 00000000 B9 oooocooo 18 oooocooo 19 00000020 00000020 17050201 00000000 Mi0 00000000 B10 00000800 11
13. 1073 0 06 0 18 0 24 Volts The transient analysis was carried out only for the frequency of 1KHz The characteristic graph is shown in figure 15 below U output U innput Time Figure 15 This is the characteristic graph generated by PSPICE transient analysis of the op amp circuit The green line corresponts to the output signal while the red to the input signal MSc Project Report 37 Echo cancellation using ADSP 21065L 9 0 1 2 3 4 5 6 7 8 9 10 References Kondoz M A 1994 Digital Speech Coding for Low Bit Rate Communication Systems page 330 335 Surrey John Wiley amp Sons Owens J F 1993 Signal Processing of Speech London The Macmillan Press LTD Analog Devices 1999 ADSP 21065L EZ LAB Development System Manual page 1 6 Shrivenham technology International Europe Limited http products analog com products_html list_gen_157_21 html Analog Devices 1999 Getting Started guide Norwood USA Analog Devices Inc Computer Products Division One technology Way http www analog com support Design_Support html http www analog com techsupt application_notes AD1819A_ 21065L pdf http www analog com techsupt software dsp app_note 21k_books html Installation path of VisualDSP 21k EZ KITs 21065L Demos Tt dpj Installation path of VisualDSP 21k EZ KITs 21065L Demos MSc Project Report 38 Echo cancellation using ADSP 21065L 10 0 Bibliography
14. A OS O oom ia D 3 CHI 240m CH1 1 00Y CH2 1 004 M Am Figure 23 Actual system output minus ideal system output MSc Project Report 48 Echo cancellation using ADSP 21065L APPENDIX II Program Code This appendix contains the assembly code necessary to create an echo canceller using the Analog Devices ADSP 21065L Digital Signal Processor It was written compiled and executed with VisualDSP software under windows 98 environment The files included in the order presented are e new65Ldefs h e def210651 h e EZLAB 21065L_debugger ldf e Lms asm e Talkthru asm e SDRAM Init asm e Clear _SPTI_regs asm e AD1819a_initialization asm e Codec Processing ISR asm e Inn 065L_EZLAB asm e ISR table asm These files along with examples used and the report itself are included in the CD ROM provided with the report The following page has a couple of pictures of the DSP board used in the project MSc Project Report 49 Echo cancellation using ADSP 21065L Custom made cables Left Right Channel In Left right Channel Out Board under test z annen Sound CODEC DSP Chip Line Out Line Mic In COM2 port MSc Project Report 50 Echo cancellation using ADSP 21065L
15. CONTENTS CONTENT Ss EE E ES 1 ACKNOWLEDGEMENT Sivad di UN 3 ABS LRA CT nta Ra ARA E Yo 4 LE INTRO KICHE Eeer 5 1 1 Chapter REN E E in E 5 1 2 Profeco IN E 5 1 3 Project Aims Tee Deele 5 1 4 o AAA O Ee 6 20 BASICTHEORY OF ECHO cursis Arda 7 2l Chapter Overview cyt ola di e GE 7 22 Whats Echo sai at 7 2 2 1 Acoustic Echo EE o a 7 2 2 2 MI A ae at aoe E 8 23 Echo Cancellation sorset N A aan ia a a a s 8 2 3 1 Echo Stppressets bcn 9 232 Adaptive Filter Echo Camco nl ef asalta ias 9 30 SHARD WARE E Ls cartes A E 12 3 1 Chapter Overview ce acc haa EEN 12 3 2 Analog Devices AWS 2106 EE 12 3 2 1 NEE E 13 3 2 2 The AJOC OD A os E EG 14 3 2 3 ADSP NOOSESHARC Key Features ococcoccconnonnnnncnncocnnonacancncnncnoncos 14 40 SOFTWARE ia ai 15 4 1 Chamet OVGRIOW nissan 15 AD VERDS En OME ri aia 15 50 AAA Eet 19 5 1 Opt BEEN 19 IA KA daptive EE 19 5 3 Transversal Pl 19 Sh Adaptive Filter s Algorithms dee edE TEEN NSdEedO Sinead dadas daros 21 5 4 1 Mathematical Mogdellmg viii aii ci ENNEN 21 MSc Project Report 1 Echo cancellation using ADSP 21065L 6 0 The Adaptive Echo Canceller ui task awenions 24 6 1 Chapter Oven Wisin eet eet det geheier Eed eet 24 6 2 Hirst To Final Step snes ee eege 24 6 2 1 The Talk Through Example A dah grime 24 6 2 2 Fixed Coefficients FIR Eterna 24 6 2 3 he IN OIse E E d E E 26 System Process BEE coeds 32 TO CONCLUSIONS eech Besse cita 34 8 0 FURTHER DEVELOPMENT std di EE 35 90
16. Process Overview CODEC SPORT DSP Initialisation S y n TMS Routine execution d n e n d n Y n System output e n MSc Project Report Echo cancellation using ADSP 21065L 32 Further details can be found in APPENDIX II in the program source code The author believes that the general concept of the design is more important than the code since somebody else might try to implement this system with another language As it is known for example the ADSP 21065L in combination with VisualDSP support the development of applications written in ANSI C MSc Project Report 33 Echo cancellation using ADSP 21065L 7 0 CONCLUSIONS The echo canceller system developed in this project and presented in this report proved to work satisfactorily and exceptionally good as a noise canceller as well The noise canceller was the first system developed since the echo can be considered as a case of noise with spectrum characteristics close to those of the echo creator signal The system utilized the Analog Devices ADSP 21065L EZ LAB development board with the ADSP 21065L DSP processor on board and the AD1819a sound CODEC The CODEC provided for dual channel input and output at 16 bits precision The echo canceller was based on a simple and reliable system design ofan adaptive filter that predicted the echo in the input signal so as to perform a subtraction of the prediction from the input signal to provide a clear output signal The ada
17. The Processor A powerful 32 bit processing 32 bit words essential for processing 20 and 24 bit input signals It is capable of 60 MIPS Million Instructions Per Second with an average 180 MOPS Million Operations Per Second and 198 MFLOPS Million Floating Point operations per Second It executes one instruction per cycle from a 16K 32 bit 544 Kbits of user configurable on chip memory that reduces memory access bottlenecks and improves processing speed It also incorporates two Data Address Generators DAGS an advanced program sequencer one result Bus and three arithmetic units ALU MAC Shifter It works at 3 3Wolts Figure 5 below illustrates the internal processor architecture for the ADSP 2100 family processor in which ADSP 21065L is included Ref 2 3 4 Data Address Data Address Program Generator Generator Sequen Output Reg Output Reg Output Reg 16 Results BUS Figure 5 ADSP 2100 Family Base Internal Architecture MSc Project Report 13 Echo cancellation using ADSP 21065L 3 2 2 The Audio CODEC The developing board includes an onboard full duplex 16 bit stereo audio CODEC the AD1819A It supports sampling frequencies of up to 48KHz and has two channel input and two channel output ports with an option at input to choose between Line Input and Microphone Input 3 2 3 ADSP 21065L SHARC Key Features The following bullet point sum up the key features of the hardware used Ref 2 3 198 MFLOPS 32 bit floatin
18. a lot of time to calculate and update for a DSP such as the one utilized in this project To overcome such a problem an echo delay estimator may prove critical since it could start the process of adaptation and echo cancellation from the moment the echo signal has almost reached the near end so that 1t is cancelled as if its delay was smaller e Anamplifier for use of microphones with the board MSc Project Report 35 Echo cancellation using ADSP 21065L An amplifier with a gain of 10 is necessary for using microphones with the board since an odd thing happened when the on board amplification capability of the board was used Only the right input channel was amplified So a simple circuit is provided as an external microphone amplifier The below schematic was designed and tested in PSPICE and is an operational amplifier with a gain of 10 e 100n 100k gt gt 900k Figure 14 Schematic of an operational amplifier with a gain of 10 Op amp circuit We want a gain of 10 therefore we have 10 Vo Vi t R2 R1 gt gt R2 gt 9R so if we choose R1 to be 100K then R2 must be 900K The values of R4 and R5 220K to split the voltage from 12 to 6 Volts For calculating the values of the capacitors we use the following formula f 2RC 1 Total offset voltage MSc Project Report 36 Echo cancellation using ADSP 21065L Vo offset VIO RI R2 R1 HOR2 6x10 3 1000x103 100x1073 200x10 9x900x
19. annel 2 the same sine wave with the added noise d n This signal was used as the system input d n MSc Project Report 42 Echo cancellation using ADSP 21065L The following two graphs show the system output in contrast with the system input The noisy system input is a perfect sinusoidal waveform signal from a signal generator plus a white Gaussian noise signal from a noise generator Channel Y top part of graph shows the noisy input signal d n while channel 2 bottom part of graph shows the system output e n d n Y n The cancellation of noise is obvious Ideally the system output should have been a perfect sinusoidal wave The difference between the graphs is in the time scale used 250us for the first 2 50ms for the second DISPLAY H Se Ms ss AM Persist A LSW DH Ganz AA PO Contrast Increase H d Y NM s E ee easy ee E ges ontrast CH2 1 004 M 250us CHI 240mY CH1 1 00 Figure 18 Chamnel 1 displays the noisy system input d n while channel 2 displays the system output e n Ideally this should be a pure sine wave MSc Project Report 43 Echo cancellation using ADSP 21065L H Tek d Trig d M Pos 1 250m MEASURE CH1 D Di Freq 1 596kH2 H 1 1 A wm E t FEEL REKE SO SEREEN EAN UMBRELLA 1 523kHz s s krpa ee O AEE ENEF EES TAASEN E eye EH h y a i UN i y t D s y
20. ce mismatch MSc Project Report 5 Echo cancellation using ADSP 21065L The echo canceller would have to utilize a DSP Digital Signal Processor for processing the necessary adaptive filtering algorithm The DSP used was the ADSP 21065L SHARC The algorithm was the well known and established LMS Least Mean Square To achieve the objective the work was roughly split into the following three stages STAGE 1 Derive a system design and figure its mathematical modelling This was also the first aim STAGE 2 Set up the DSP hardware and start the implementation of the theoretical design STAGE 3 Test and debug the system to achieve real time echo cancellation 1 4 Equipment Used Below is a list of the equpment used to create and test the echo canceller system e Signal generatorl LEVELL Function Generator Type TG 303 e Signal generator2 LEVELL Function Generator Type TG 303 e DSP Chip ADSP 21065L EZ LITE Developing Board e Computer Software PC running windows98 with Analog Devices VisualDSP software installed e Oscilloscopel HAMEG Storage Oscilloscope HM 205 3 e Oscilloscope2 Digital Oscilloscope used for generating printout results e S t of custom audio Cables e Noise Generator H01 3722A Noise Generator MSc Project Report 6 Echo cancellation using ADSP 21065L 2 0 BASIC THEORY OF ECHO 2 1 Chapter Overview This chapter describes basic information about the types of echo in communication systems
21. cellation using ADSP 21065L 18 5 0 Adaptive Filtering 5 1 Chapter Overview In this chapter the author explains some fundamental concepts governing adapting filters that is relevant to the echo canceller implementation project undertaken 5 2 Adaptive Filters As mentioned in the end of chapter 2 an adaptive filter is a digital filter whose characteristics change in an unknown environment input signal Whether the adaptive filter is used in communication sonar radar or image signal processing the basic feature remains the same A set of adjustable coefficients are changed through the error estimation performed by an algorithm that has an input vector with a desired response From this we can deduce the two basic parts of an adaptive filter e Adjustable coefficients e Algorithm that updates adjusts the coefficients There are many types of digital filters each performing best on specific tasks yet they are categorized either as Finite Impulse Response FIR filters that as implied by the name have a finite length impulse response or Infinite Impulse Response IIR or recursive filters that has an impulse response of infinite length Both types of filters can be used in the realization of the echo canceller but an FIR filter was chosen since it is simple and stable 5 3 Transversal FIR In the echo canceller implementation a transversal or tapped delay line FIR filter was used since it is relatively easy to be implement
22. ed and provides stable and accurate performance The diagram below Fig 9 shows the realization of the filter MSc Project Report 19 Echo cancellation using ADSP 21065L x n N 1 Figure 9 A Transversal FIR filter implementation The squares with Z in them denote the delay elements of the filter The ovals with the W n starting at zero up to N 1 correspond to the weights of the filter or else coefficients The circles are adders The above diagram corresponds to the following difference equation Ref 5 N I Y n W n X n Y W n X n i equation 2 i Where X n is the input vector and W n the weights vector N is the number of filter taps coefficients which give the length of its impulse response For a sampling frequency of 8kHz the number of filter taps give the total echo path the system can process Typically for N 64 an echo delay of 8ms can be fully processed The diagram on the next page Fig 10 illustrates the position of the echo canceller unit at the far end and near end positions in a communications system In the project undertaken the number of taps utilized was 64 as a trade of between echo path and rapid filter convergence when turned on In reality the filter performs satisfactorily in echo paths of up to 20ms MSc Project Report 20 Echo cancellation using ADSP 21065L Near end Far End Speaker Listener Comm Channel Near End Far End Listener Speaker Figure 10 A basic illustration of echo
23. f 1Volt at the Right Channel of the CODEC input A white noise signal generator was connected to the Left Channel of the CODEC input providing the noise into the system with maximum amplitude of 0 3Volts A second signal generator instead of the noise generator was used while testing the system for echo cancellation response An oscilloscope was connected to the CODEC Line out port to monitor for the two 2 outputs The talk through example program used 48KHz sampling frequency This frequency was changed to 8kHz by modifying the necessary variable in the AD1819a initialization asm file This change in sampling frequency was done because the maximum frequency transmitted in voice telephone networks is 3 5KHz so a sampling frequency of 8Khz is adequate MSc Project Report Dd Echo cancellation using ADSP 21065L The Line In port was preferred to that of the Microphone In by also changing the necessary value in the AD1819a_initialization asm file and also by setting the necessary jumpers on the developing board The jumper pins are a set Tor each channel Left Right as shown in figure 12 below E Line In 0 Mic In Line In e Mic In Figure 12 The schematic above gives a picture of how he jumpers and pins look when selecting the Line Input The square boxes above depict the plastic sorting jumpers provided with the board The task was to display the noisy input signal op one channel and the cleared output signal o
24. g point 180 MOPS 32 bit fixed point 16K 32 bit Dual ported on chip memory 544 KBits configurable 64M x 32 bit word external address space Glueless SDRAM interface 2 serial transmit receive ports Tx Rx support 32 channel TDM DS mode supports up to 8 channels Two timers with event capture and PWM options 12 programmable I O pins 10 DMA channels Glueless 32 bit SDRAM Interface multiprocessing with 2 ADSP 21065Ls with synchronous data transfer rate 3 3Volt 208 pin PQFP TDM interface Direct interface to T1 E1 lines 240 Mbytes second external port Maximum data throughput Non inttusive DMA Utilize full capabilities of core processor MSc Project Report 14 Echo cancellation using ADSP 21065L 4 0 SOFTWARE 4 1 Chapter Overview The following paragraphs describe the software used to write the code of the echo canceller The author believes that a detailed analysis of the steps to install and run the software is beyond the scope of this report and suggests further reading from the installation and software manuals included in the package Ref 2 4 4 2 VisualDSP Environment In the ADSP 21065L SHARC EZ LAB a CD ROM was included containing a software program for developing DSP applications The program s name was Visual DSP and is comprised of an Integrated Development Environment and a Debugger The environment provided a project management utility containing some useful tools such as a C compiler a run time library with
25. ignal we want to filter d n At the same time the filter takes just the noise x n and by performing the computations gives its prediction Y n MSc Project Report 29 Echo cancellation using ADSP 21065L r0 31 rl 31 f0 float r3 by 10 fl float r9 by rl dm i0 m0 f0 f4 pm i8 m8 store x n in delay line f4 w0 n f8 f0 f4 fO dm i0 m0 f4 pm i8 m8 f8 x n w0 n f0 x n 1 f4 w1 n f12 f0 f4 f0 dm 10 m0 f4 pm i8 m8 f12 x n 1 w1 n f0 x n 2 f4 w2 n Icntr T APS 3 do macs until Ice macs f12 f0 f4 f8 f8 f12 f0 dm 10 m0 f4 pm 18 m8 f12 x n i wi n f8 sum of prod f0 x n i 1 f4 wi 1 n f12 f0 f4 f8 f8 f12 f12 x n N 1 wN 1 n f13 f8 f12 f13 y n The brackets contain the comments that explain the individual code line This prediction is subtracted by the noisy signal d n and this yields the prediction error and system output e n e n d n y n equation 1 fo fl f13 D e n Where floating point register fl holds the noisy signal sample from register r9 This error is processed by the LMS routine to do the necessary updating of the coefficients Wi n Wi n h Wi n 2ue n x n i Where 2u is the value of the STEPSIZE variable defineSTEPSIZE 0 005 MSc Project Report 30 Echo cancellation using ADSP 21065L f1 f6 f7 f4 dm i0 m0 f1 STEPSIZE e n f4 x n f0 f1 f4 f12 pm i8 m8 f0 STEPSIZE e
26. led the user to input a voice signal by means of a microphone to the stereo input port of the developing board and listen to an output on speakers connected to the stereo line out port of the developing board After a lot of experimentation with changing the code a better understanding of the procedures involved gradually was developed 6 2 2 Fixed Coefficients FIR Filter As a first target the implementation of a simple low FIR digital filter with fixed coefficients was set Below is part of the implementation of the filter in assembly code given bere as a code example Ref 6 9 The filter loaded the coefficients from a text file into a circular buffer took the samples of the input signal form the CODEC via a register 12 and execute the filtering routine to provide an output of the form MSc Project Report 24 Echo cancellation using ADSP 21065L N I Y n W n X n X Walin i Finite Impulse Response Filter FIR r12 r12 xor r12 dm 11 0 12 set r12 0 store input sample in line circular buffer r8 r8 xor r8 fO dm 10 0 f4 pm 18 m8 set r8 0 get data get coefficients from buffer lentr FIRLen 1 do macloop until Ice macloop 112 f0 f4 f8 f8 f12 f0 dm 11 m1 f4 pm 18 m8 MAC rts db return from delayed subroutine f12 f0 f4 f8 f8 f12 Aast multiplication f0 f8 f12 Aast accumulation The coefficients for the filter were calculated with the help of a program EZ FIR that a
27. loscope MSc Project Report 40 Echo cancellation using ADSP 21065L The following graph displays the predicted noise signal corresponding to the adaptive filter output Y n on channel 1 of the oscilloscope and the noise signal x n from the noise generator on channel 2 of the oscilloscope yasan aa a 1 wm D ee IRA MO s d R Pe cae K s H s i t sey CO S A gt e oN gS R D e z M A Nas Da AA AS La sae 1 Y e B H H H HM H rae De s 4 S508 Irmy A AO CS ACT eee ee ee CHI 1004 CH2 1004 M 250us Figure 16 Channel 1 displays the predicted noise Y n Channel 2 displays the actual noise x n MSc Project Report 41 Echo cancellation using ADSP 21065L The following graph displays the pure sinusoidal signal from the signal generator on channel 1 of the oscilloscope while on channel 2 of the oscilloscope the noisy noise sine wave system input d n Tek Ji Trig d M Pos 1 220ms MEASURE y APAR SS Kee KAISER IRAK CR Me O O Ree KREIEN H H H a H H s pal ee AM GOR PO E O dE WE CH x 8 s Freq KK W 1 52anHe e EE ss AMES bx o A ZY FEEF AIIAEAKEEKEKEKERIIIAEAI w wt is D SA 1 37 6KH2 See 0 ws y MA ss ss sos Mean 32 3mV sv ERA ss ss A y ss ss sy sr AR ss ss Mean N Gm CHT 7 b20m Ven Figure 17 Channel 1 displays the sine wave from the signal generator and ch
28. n x n f12 w0 n Icntr TAPS 1 do update_weights until Ice f8 f0 f12 f4 dm i0 m0 f12 pm i8 m8 f8 wi n 1 f4 x n i 1 f12 wit 1 n update weights f0 f1 f4 pm i9 m8 f8 f0 STEPSIZE e n x n i 1 store wi n 1 rts db f8 f0 f12 f0 dm 10 1 f8 wN 1 n 1 10 gt x n 1 location in delay line pm i9 m8 f8 store wN 1 n 1 The whole procedure as obvious from studying the code repeats itself every time a new set of samples arrives from the receive port of the CODEC Circular Buffering Circular buffers are nothing more than normal buffers with the advantage that after reaching the end of its length it jumps back to the beginning holding one by one the new values overwriting the existing ones automatically This kind of buffer was used to hold the input data as well as the filter coefficients The data buffer used was deline_data The coefficient buffer was weights DAG 1 was used to generate the delay line data buffer and DAG 2 to create the filter weights buffer The initialisation of the filter and algorithm was as follows lms init bO deline_data m0 1 IO TAPS circular delay line buffer b8 weights b9 b8 m8 1 18 TAPS circular weight buffer MSc Project Report 31 Echo cancellation using ADSP 21065L 19 18 f7 STEPSIZE f0 0 0 Icntr TAPS do clear_bufs until Ice clear_bufs dm i0 m0 f0 pm i8 m8 f0 clear delay line amp weights rts System
29. n the other So the Left Channel of the Line Out was to output the system output and the Right channel of the Line Out was to output the input signal with the noise so that a comparison of those signals on the display of the oscilloscope could give an immediate impression whether the system works or not For easy access to the input and output channels and for provision to microphone use the use of six male and four female stereo jacks was devised A cable used in stereo audio equipment such as the one used for stereo headphones was split in half so as to produce the following accessory illustrated in figure 13 MSc Project Report 28 Echo cancellation using ADSP 21065L To Line In To Line Out Cleared system output Right Channel Signal noise output nel OS Signal from noise generator noise Right Channel Signal from signal generator speaker Figure 13 An illustration of the custom made cables used in the project The noise x n plus the signal was generated by adding the clear signal to the noise signal with code r5 dm Right Channel In r3 DM Left Channel In noise 19 r5 13 DM Right Channel Out r9 Register 5 holds the sample coming in from the CODEC s right channel Register 3 holds the sample coming in from the CODEC s left channel Register 9 holds the noisy sample CODEC s right channel output is the contents of register 9 This way the oscilloscope displays the noisy s
30. ng ADSP 21065L 2 3 1 Echo Suppresser Echo suppresser is mainly a device that detects the level of speech on an active channel Depending on the threshold of that level anything above is treated as speech and passes through as opposed to anything below that is treated as noise or echo and is therefore heavily attenuated Such a scheme may work satisfactorily for terrestrial circuits and echo delays of less than 100ms but for satellite communication were 400ms echo delays are typical performs poorly Another problem with such a system is that it is heavily depended on the speech echo classification system it involves This sub system discriminates between forward speech and echo based mainly on the signal amplitude level high level for speech low level for echo and is exactly this feature that poses a problem when sometimes it allows high level of echo to pass through as speech and low level of speech to be attenuated as echo Furthermore an echo suppressor cannot operate during double talk and produces choppy echo These are also the main reasons why echo suppressors are being replaced by adaptive echo cancellers 2 3 2 Adaptive Filter Echo Canceller As mentioned in the introduction adaptive echo cancellers were initially developed during the 1960 s by AT amp T Bell laboratories The implementation of an adapting echo canceller system was introduced in a satellite network where echo delays exceeded 600ms The project was a success As illus
31. ptimization Testing its efficiency and omitting parts that may not be necessary could optimise the code of the system That is because many core part were imported from examples and applications of different nature e Active channel detection A key feature of an AEC algorithm is active channel detection Indeed when the far end operator is silent and the near end operator is talking the filter must not be adapted because the near end operator is no longer an echo e Double Talk Detection During a DT condition the near end signal on the microphone contains echo and near end speech The residual error used to update the filter coefficients includes the near end speech and if the algorithm is still adapting the algorithm may start to diverge To cure this problem the filter coefficients must not be updated during that time e Use of another more efficient algorithm The use of an algorithm such as the normalized least mean square NLMS for example could improve the system e Echo Delay Estimator The echo canceller developed in this project task was not tested to verify its ability to cancel echo that results from long distance communications such as in a satellite system In such a system echo paths of 500ms are typical and some times exceed even 800ms For an echo path of 500ms to be represented by the echo path length of an adaptive filter 4000 taps are needed if sampling is performed at 8KHz This large number of filter coefficients takes up
32. ptive filter that was used in this project was a transversal FIR Finite Impulse Response filter with the LMS Least Mean Square algorithm handling the adaptation of the filter coefficients This type of popular filter was chosen for its simplicity and stability and because it is widely used on adaptive filter designs for sound applications world wide The algorithm was chosen to be the LMS since it is popular among the choices of entry level adaptive filter designers and because it is easy to implement and monitor Its performance may not be the best but is satisfactory The NLMS Normalized Least Mean Square was considered as a second algorithm implementation but due to time restrictions the thought was discarded One aspect of the eche canceller that failed to reach expectations was its use in satellite communications This is due to the echo path length in such a system that exceeds 400ms Some solutions to this problem are mentioned in chapter 8 The echo canceller systeman its current form cancels noise exceptionally good and echo with up to 10ms delay APPENDIX Hincludes the result graphs obtained while monitoring the performance of the system and APPENDIX II the program source code documented MSc Project Report 34 Echo cancellation using ADSP 21065L 8 0 FURTHER DEVELOPMENT There are numerous activities that could better the performance of the current echo canceller The most important in the authors opinion are e Code o
33. r wire trunk network Two wire Subscriber Subscriber Loop S1 Loop S2 OQ OO N ZN Figure 2 Echo Transmission within a telephone hybrid network The hybrids function is to direct the signal energy from S1 and S2 to the upper and lower paths of the four wire network circuit without allowing any leakage back to the two sources over the opposite direction line pairs Ref 1 Unfortunately due to the impedance mismatching some of the transmitted signal returns to its original source thus an echo occurs A delay of up to 1 2 milliseconds can be tolerated or even go unnoticed but as the signal round trip increases the echo becomes more noticeable and annoying Typical values for long distance communication satellite communication round trip is 300 600 msec delay This proves to be a serious problem 2 3 Echo Cancellation There are various techniques for cancelling echo The two most prevailing in everyday modern communication systems are Echo Suppression and Adaptive Filtering Echo suppression the oldest technique for echo cancellation dates back to the 1950 s It was used both in terrestrial and non terrestrial communications but exhibited poor performance in the later Adaptive filtering was first introduced in 1960 s by AT amp T Bell laboratories and progressively became more and more popular as digital technology allowed compact fast and economic real time signal processing units DSP s MSc Project Report 8 Echo cancellation usi
34. sal FIR filter with LMS algorithm updating the coefficients The algorithm although not the most efficient of those existing nowadays is widely used basically due to the fact that needs a small amount of memory is relatively simple to implement and appears to be stable without providing any care The core assembly for the task was found after some research in a file at the site of Analog Deviees under the name lms rls zip Ref 7 This file contained several assembly codes for testing well known adaptive filter algorithms The initial file used in the project was called Ims asm and was written by Wassim G Najm from Analog Devices DSP division This code though needed modification to be incorporated to the already changed talk through example program Since echo could be regarded as a case of having noise in a signal the system was biased towards an exhibition of noise cancelling properties For this purpose the set up of the equipment was as illustrated in figure 11 MSc Project Report 26 Echo cancellation using ADSP 21065L White Noise Generator Signal Generator Right Channel Computer ADSP21065L VisualDSP Developing Board Environment Oscilloscope Left Channel Right Channel Channel 2 Figure 11 General overview of instruments and connections between them while testing the system A signal generator was used to provide an input signal with a sinusoidal frequency of 2KHz and peak to peak amplitude o
35. t so close to the optimal ones therefore the system performs fast but not accurately By decreasing the value ofu the weights change slower but their values get closer to the optimal ones This wan the system converges slowly but is accurate Typically 0 lt u lt 1 and for this implementation was 0 0025 The true gradient V n can be estimated by the em to be an estimate of the mean square error Thus which if substituted in equation 7 yields the LMS algorithm w n 1 w n 2ue n x n equation 8 MSc Project Report 23 Echo cancellation using ADSP 21065L 6 0 The Adaptive Echo Canceller 6 1 Chapter Overview In this chapter the actual implementation of the system is described by presenting the steps followed to achieve the end result of a working application The code is not explained in detail since such a thing would be pointless Only key points are addressed The code provided in the appendix is full and the comments within are mostly self explanatory 6 2 First To Final Step 6 2 1 The Talk Through Example Due to the authors lack of previous DSP application developments he deemed necessary to make a start from the very basic demo programs included in the Visual DSP software package So a ready written project called Talk Through was used as the basis for the development of the echo canceller This project involved some files and a very simple program that initialised the sound CODEC on the developing board and enab
36. trated in figure 3a b below the basic function of an adaptive echo canceller is to predict the echo Y n added to the receive path signal dn and subtract it thus giving a clear output signal e n This echo cancelling systems appropriate for both types of echo namely electric fig 3a and acoustic fig 3b MSc Project Report 9 Echo cancellation using ADSP 21065L Far End x n Speaker Two wire Subscriber Adaptive een Filter OO Los Far End Listener e n d n Figure 3a The adaptive filter removes Far End Speaker echo from hybrid return path d n The location of the filter is as close to the Far End Speaker Listener as possible Far End x n Near End Speaker Q A Listener Adaptive Filter Near End Far End O Speaker Listener e n d n Figure 3b The adaptive filter removes Far End Speaker echo from the Near End Speaker path dn The location of the filter is as close to the Far End Speaker Listener as possible The adaptive filter is a digital filter with variable coefficients which means its frequency response can be modified The coefficients update according to some criterion to give the desired prediction for the echo signal by means of an algorithm that continually monitors the error output of the system More specifically the predicted echo is generated by the input signal x n and the system error e n en d n Y n equation 1 MSc Project Report 10 Echo cancellation using ADSP 21065L
37. utomatically generated the desired filter weights by choosing from the following information e Filter Type Low pass High pass Band pass Band stop Hilbert transformer Differentiator e Window Type Rectangular Triangular Hanning Hamming Generalized Hamming Kaiser Bessel e Filter Length 1 to 256 e Sampling Frequency e Cut off Frequency e Filter Gain MSc Project Report 25 Echo cancellation using ADSP 21065L By using the EZ FIR program and entering the following specifications Low pass filter Rectangular window 64 taps filter length 8000Hz sampling frequency 3 5KHz cut off frequency and a gain of 1 a set of coefficients was generated and loaded to the buffer with index 18 The program was compiled linked loaded run and monitored from the debugger To see though if it really worked a test had to be performed The filter was tested using a signal generator providing the input signal at the line in mode of the board and an oscilloscope monitoring the line out of the board The filter performed as expected with a deviation of approximately 400Hz below the specified cut off value A number of other filter types and specifications were also designed and implemented with the same FIR structure for the author to be even more familiarized with the code and DSP board 6 2 3 The Noise Canceller As the next stage of a successfully working fixed coefficients FIR filter it was decided to move on and implement a transver
38. where e n is the system output d n is the incoming signal with the actual echo and Y n is the predicted echo There are many types of adaptive filter designs used in a variety of applications where frequency manipulation of a signal is needed In this project a transversal FIR Finite Impulse Response filter was used and is discussed later on chapter 5 3 MSc Project Report 11 Echo cancellation using ADSP 21065L 3 0 HARDWARE 3 1 Chapter Overview This chapter gives information and basically presents the hardware used in this project Namely the ADSP 21065L SHARC digital signal processor 3 2 Analog Devices ADSP21065L Analog Devices is an American company that specializes in designing and manufacturing a variety of linear mixed signal and digital signal processing integrated circuits ADSP 21065L is a member of a large family of DSP chips utilising the Super Harvard Architecture SHARC Its a 32 bit general purpose DSP that allows programming in fixed or floating point arithmetic The chip comes with a developing board under the package named ADSP 21065L SHARC EZ LAB The following figure Fig 4 illustrates the general structure of the device CLK i 30 MHz 120 Mbytes sec TU i a 240 Mbytes sec SDRAM ed 420 Mbytes sec ts 30 Mbits sec ve Optional Mulfi Processing Shown Figure 4 General Overview of ADSP 21065L MSc Project Report 12 Echo cancellation using ADSP 21065L 3 2 1
Download Pdf Manuals
Related Search