|
5
|
1 Open Source Line Echo Canceller (OSLEC)
|
|
|
2 =======================================
|
|
|
3
|
|
|
4 [[introduction]]
|
|
|
5 Introduction
|
|
|
6 ------------
|
|
|
7
|
|
|
8 Oslec is an open source high performance line echo canceller. When
|
|
|
9 used with Asterisk it works well on lines where the built-in Zaptel
|
|
|
10 echo canceller fails. No tweaks like rxgain/txgain or fxotrain are
|
|
|
11 required. Oslec is supplied as GPL licensed C source code and is free
|
|
|
12 as in speech.
|
|
|
13
|
|
|
14 Oslec partially complies with the G168 standard and runs in real time
|
|
|
15 on x86 and Blackfin platforms. Patches exist to allow any Zaptel
|
|
|
16 compatible hardware to use Oslec. It has been successfully tested on
|
|
|
17 hardware ranging from single port X100P FXO cards to dual E1 systems.
|
|
|
18 Hardware tested includes TDM400, X100P, Sangoma A104, Rhino E1 etc.
|
|
|
19 It also works well on the link:ip04.html[IP04 embedded Asterisk
|
|
|
20 platform].
|
|
|
21
|
|
|
22 There is also a project underway to use
|
|
|
23 http://peter.schlaile.de/mISDN/[OSLEC with mISDN].
|
|
|
24
|
|
|
25 Oslec is included in many distributions, including Debian, Gentoo,
|
|
|
26 Trixbox, Elastix, and Callweaver.
|
|
|
27
|
|
|
28 [[testimonials]]
|
|
|
29 Testimonials
|
|
|
30 ------------
|
|
|
31
|
|
|
32 So how good is Oslec?
|
|
|
33
|
|
|
34 + The http://www.trixbox[Trixbox] community have been testing
|
|
|
35 Oslec, here is a
|
|
|
36 http://www.trixbox.org/forums/trixbox-forums/open-discussion/need-people-echo-problems[Trixbox
|
|
|
37 forum thread] that has feedback from many people. Start about half
|
|
|
38 way down the page, or search on the string "night and day" :-)
|
|
|
39
|
|
|
40 + The http://callweaver.org/wiki/Zaptel[CallWeaver] project use and recommend Oslec.
|
|
|
41
|
|
|
42 + Here is some feedback from Oslec users:
|
|
|
43
|
|
|
44 [Gordon Henderson, posted in the Asterisk-user mailing list July 21 2008]
|
|
|
45
|
|
|
46 ________________________________________________________________________
|
|
|
47 I switched to OSLEC after testing HPEC on TDM400 boards, and found that it
|
|
|
48 worked much better and wasn't limited to the restricted mechanism Digium
|
|
|
49 uses for licensing (unlikely as it sounds, I have some clients who do not
|
|
|
50 have a connection to the public Internet, and never will for their phone
|
|
|
51 system)
|
|
|
52
|
|
|
53 It also passes the wife test which HPEC didn't.
|
|
|
54
|
|
|
55 It's also free (OS as in Open Source), which HPEC isn't, although that
|
|
|
56 wasn't my primary reason for using it - ease of use and "workability" was.
|
|
|
57 _____________________________________________________________________
|
|
|
58
|
|
|
59 [Michael Gernoth, Author of the Zaptel MG2 echo canceller]
|
|
|
60
|
|
|
61 ________________________________________________________________________
|
|
|
62 Thanks for OSLEC, it's so much better than my hacked KB1 which is called
|
|
|
63 MG2 :-)
|
|
|
64 _____________________________________________________________________
|
|
|
65
|
|
|
66 [David Gottschalk]
|
|
|
67
|
|
|
68 ________________________________________________________________________
|
|
|
69 Once I managed to get OSLEC installed (it was a big job for me, because I was
|
|
|
70 missing many of the dependancies and knowledge), it immediately fixed the echo
|
|
|
71 problems I had been unable to get rid of to date despite lots and lots of
|
|
|
72 tweeks, tests and fiddling. Even more suprisingly, it appears to converge
|
|
|
73 almost immediately from my initial test.
|
|
|
74
|
|
|
75 This code is the best since thing since, well, Asterisk !!!
|
|
|
76 _____________________________________________________________________
|
|
|
77
|
|
|
78 [Nic Bellamy, Head Of Engineering; Vadacom Ltd]
|
|
|
79
|
|
|
80 ________________________________________________________________________
|
|
|
81 Short version: KB1 < MG2 < OSLEC < HPEC
|
|
|
82
|
|
|
83 In the open source line echo canceller category, OSLEC is so far the
|
|
|
84 best I've seen.
|
|
|
85
|
|
|
86 It converges reasonably fast, and thanks to it's dual-path approach,
|
|
|
87 will converge, albeit a bit slowly, on nasty echo paths where cancellers
|
|
|
88 like KB1 and MG2 will fail entirely.
|
|
|
89
|
|
|
90 The nonlinear processor and comfort noise generator provide for a far
|
|
|
91 more pleasant listening experience than KB1/MG2.
|
|
|
92
|
|
|
93 It's not quite up to the grade of ADT's G.168 canceller used in Digiums
|
|
|
94 HPEC product in terms of convergence speed or tail coverage.
|
|
|
95 _____________________________________________________________________
|
|
|
96
|
|
|
97 [Pawel Pastuszak, Astfin.org]
|
|
|
98
|
|
|
99 ________________________________________________________________________
|
|
|
100 Since I installed the first version of OSLEC I was hooked.... It's
|
|
|
101 been several months now since the alpha version was released; after
|
|
|
102 using OSLEC Mark and I agreed that we had to have OSLEC as the main
|
|
|
103 echo cancellation in the Astfin distribution.
|
|
|
104 ________________________________________________________________________
|
|
|
105
|
|
|
106 [Pavel Selivano, Parabel Ltd]
|
|
|
107
|
|
|
108 ________________________________________________________________________
|
|
|
109 I've been working with Oslec since March of 2007. Oslec has had some
|
|
|
110 problems, but, thanks to David, they are solved. Now, I can say Oslec
|
|
|
111 is time-tested. The best, I can note - it is still under development
|
|
|
112 (with feed-back). Thank you David for your good deed.
|
|
|
113
|
|
|
114 BTW: comparison to Intel IPP's echo canceler have shown, Oslec is very
|
|
|
115 good :-)
|
|
|
116 ________________________________________________________________________
|
|
|
117
|
|
|
118
|
|
|
119 [[news]]
|
|
|
120 Oslec News
|
|
|
121 ----------
|
|
|
122
|
|
|
123 + Saturday, June 7, 2008: We have worked out how to use <<rhino,Oslec with
|
|
|
124 Rhino>> PRI cards. Thanks Juan Manuel Coronado!
|
|
|
125
|
|
|
126 + Friday, May 16, 2008: Oslec Road Map. I have kicked off a
|
|
|
127 discussion of the
|
|
|
128 http://sourceforge.net/mailarchive/forum.php?thread_name=1210895745.7127.60.camel%40localhost&forum_name=freetel-oslec[development
|
|
|
129 roadmap] for Oslec. Is hardware echo cancellation headed for
|
|
|
130 extinction?
|
|
|
131
|
|
|
132 + Friday, May 16, 2008: Oslec chosen as the **default**
|
|
|
133 http://sourceforge.net/mailarchive/forum.php?thread_name=20080515131239.GJ27523%40xorcom.com&forum_name=freetel-oslec[echo
|
|
|
134 canceler for Debian]. Thanks Tzafrir for your work on packaging Oslec
|
|
|
135 for Debian!
|
|
|
136
|
|
|
137 + Tuesday, May 13, 2008: link:/blog/?p=64[Oslec versus the SPA 3000]
|
|
|
138 ATA. A blog post comparing the SPA 3000 echo cancellation to Oslec on
|
|
|
139 a problem FXO line.
|
|
|
140
|
|
|
141 + Wednesday, February 11, 2008: Oslec mailing list created. Due to
|
|
|
142 popular demand (well, Tzafrir mainly!) there is now an
|
|
|
143 https://lists.sourceforge.net/lists/listinfo/freetel-oslec[Oslec
|
|
|
144 mailing list].
|
|
|
145
|
|
|
146 + Sunday, February 10, 2008: Zaptel 1.4.8 patch checked into SVN,
|
|
|
147 thanks Tzafrir.
|
|
|
148
|
|
|
149 [[install]]
|
|
|
150 Installing OSLEC with Asterisk/Zaptel
|
|
|
151 --------------------------------------
|
|
|
152
|
|
|
153 Notes:
|
|
|
154
|
|
|
155 + This process also installs a system for sampling echo signals that
|
|
|
156 is helpful for developing oslec.
|
|
|
157
|
|
|
158 + I assume Asterisk is already installed and tested.
|
|
|
159
|
|
|
160 + I assume you are running Linux 2.6 and Zaptel 1.2.13 or a later
|
|
|
161 1.2/1.4 version of Zaptel.
|
|
|
162
|
|
|
163 + I assume you are using a Digium TDM400 line interface card. Change
|
|
|
164 the "insmod wctdm.ko" line below to match your line interface
|
|
|
165 hardware (e.g. wcfxo for a X100P).
|
|
|
166
|
|
|
167 + For Zaptel 1.4.x replace the 1.2.13 text below 1.4.x (tar ball and
|
|
|
168 patch file).
|
|
|
169
|
|
|
170 + For Linux 2.4 replace the "insmod moduleXYZ.ko" lines below with
|
|
|
171 "insmod moduleXYZ.o".
|
|
|
172
|
|
|
173 1/ Download, build and install oslec:
|
|
|
174
|
|
|
175 $ cd ~
|
|
|
176 $ wget http://www.rowetel.com/ucasterisk/downloads/oslec-0.1.tar.gz
|
|
|
177 $ tar xvzf oslec-0.1.tar.gz
|
|
|
178 $ cd oslec-0.1
|
|
|
179 $ make
|
|
|
180 $ insmod kernel/oslec.ko
|
|
|
181
|
|
|
182 Optional: If you want the latest and greatest, replace the "wget" step
|
|
|
183 with:
|
|
|
184
|
|
|
185 $ svn co http://svn.astfin.org/software/oslec/trunk/ oslec
|
|
|
186
|
|
|
187 NOTE: There are patches for Zaptel 1.2.13, 1.2.18, 1.2.24, 1.4.1,
|
|
|
188 1.4.3, 1.4.4, 1.4.7.1, 1.4.8, 1.4.9.2 and 1.4.11. It's quite easy to
|
|
|
189 port to other versions, please feel free to send me a new patch should
|
|
|
190 you get Oslec working with another Zaptel version. If you can't see a
|
|
|
191 patch for your Zaptel version try
|
|
|
192 http://svn.astfin.org/software/oslec/trunk/kernel[Oslec SVN] for the
|
|
|
193 latest patches.
|
|
|
194
|
|
|
195 2/ Build, patch and install Zaptel. First obtain zaptel-1.2.13.tar.gz and:
|
|
|
196
|
|
|
197 $ tar xvzf zaptel-1.2.13.tar.gz
|
|
|
198 $ cd zaptel-1.2.13
|
|
|
199 $ ./configure
|
|
|
200 $ patch < ../oslec-0.1/kernel/zaptel-1.2.13.patch (see note 1 below)
|
|
|
201 $ cp ../oslec-0.1/kernel/dir/Module.symvers . (see note 2 below)
|
|
|
202 $ make
|
|
|
203 $ insmod zaptel.ko
|
|
|
204 $ insmod wctdm.ko
|
|
|
205 $ ./ztcfg
|
|
|
206 $ asterisk
|
|
|
207
|
|
|
208 NOTE: 1: Use -p1 option for zaptel-1.4.9.2 and above
|
|
|
209
|
|
|
210 NOTE: 2: The "cp Module.symvers" step above is optional on many systems;
|
|
|
211 it stops warnings like WARNING: "oslec_echo_can_create" [zaptel.ko]
|
|
|
212 undefined!". However on some Linux machines zaptel will not compile
|
|
|
213 without this step.
|
|
|
214
|
|
|
215 3/ These options in zapata.conf are important:
|
|
|
216
|
|
|
217 ---------------------------------------------------------------------
|
|
|
218 echocancel=yes
|
|
|
219 echocancelwhenbridged=no
|
|
|
220 ;echotraining=400
|
|
|
221 ---------------------------------------------------------------------
|
|
|
222
|
|
|
223 The *echocancelwhenbridged=no* allows faxes to pass from FXS to FXO
|
|
|
224 ports without interference from the echo canceller. This option is
|
|
|
225 important for fax signals.
|
|
|
226
|
|
|
227 Make sure *echotraining* is disabled when using Oslec - this is not
|
|
|
228 supported and if enabled will cause the channel to be silent (i.e. no
|
|
|
229 audio will pass through).
|
|
|
230
|
|
|
231 4/ The settings above have been shown to reliably cancel echo in 95%
|
|
|
232 of cases. If you can still hear echo you may have one of the rare
|
|
|
233 cases where your echo is longer than 16ms. To configure Oslec with a
|
|
|
234 32ms tail:
|
|
|
235
|
|
|
236 ---------------------------------------------------------------------
|
|
|
237 echocancel=256
|
|
|
238 echocancelwhenbridged=no
|
|
|
239 ;echotraining=400
|
|
|
240 ---------------------------------------------------------------------
|
|
|
241
|
|
|
242 Then restart Asterisk.
|
|
|
243
|
|
|
244 [[pbxinaflash]]
|
|
|
245 PBX in a Flash Install
|
|
|
246 ----------------------
|
|
|
247
|
|
|
248 Here are the
|
|
|
249 http://pbxinaflash.com/forum/showthread.php?t=100&page=3[Oslec install
|
|
|
250 procedure] for PBX in a Flash. Check the 7-24-08 post in this thread
|
|
|
251 by Alex728 for Zaptel-1.4.11 instructions. Thanks JD Austin and Alex.
|
|
|
252
|
|
|
253 Matt Keys has suggested that the PBX in a Flash install procedure also
|
|
|
254 worked well for Ubuntu server.
|
|
|
255
|
|
|
256 [[rhino]]
|
|
|
257 Rhino PRI cards
|
|
|
258 ---------------
|
|
|
259
|
|
|
260 The rhino-2.2.6 driver for Rhino PRI cards has software echo
|
|
|
261 cancellation disabled by default. This means there is no echo
|
|
|
262 cancellation unless you opt for the Rhino hardware echo cancellation
|
|
|
263 module. However the driver is easy to hack to enable software echo
|
|
|
264 cancellation, and hence Oslec.
|
|
|
265
|
|
|
266 In the rhino-2.2.6 driver, file r1t1_base.c, r1t1_receiveprep
|
|
|
267 function, around line 800:
|
|
|
268
|
|
|
269 ---------------------------------------------------------------------
|
|
|
270 static void r1t1_receiveprep(struct r1t1 *rh, int nextbuf)
|
|
|
271 {
|
|
|
272 /*
|
|
|
273 int x;
|
|
|
274
|
|
|
275 for (x=0;x<rh->span.channels;x++) {
|
|
|
276 zt_ec_chunk(&rh->chans[x], rh->chans[x].readchunk, rh->ec_chunk2[x]);
|
|
|
277 memcpy(rh->ec_chunk2[x],rh->ec_chunk1[x],ZT_CHUNKSIZE);
|
|
|
278 memcpy(rh->ec_chunk1[x],rh->chans[x].writechunk,ZT_CHUNKSIZE);
|
|
|
279 }
|
|
|
280 */
|
|
|
281 zt_receive(&rh->span);
|
|
|
282 }
|
|
|
283 ---------------------------------------------------------------------
|
|
|
284
|
|
|
285 **Solution 1:** Simply un-comment the code and rebuild the driver:
|
|
|
286
|
|
|
287 ---------------------------------------------------------------------
|
|
|
288 static void r1t1_receiveprep(struct r1t1 *rh, int nextbuf)
|
|
|
289 {
|
|
|
290
|
|
|
291 int x;
|
|
|
292
|
|
|
293 for (x=0;x<rh->span.channels;x++) {
|
|
|
294 zt_ec_chunk(&rh->chans[x], rh->chans[x].readchunk, rh->ec_chunk2[x]);
|
|
|
295 memcpy(rh->ec_chunk2[x],rh->ec_chunk1[x],ZT_CHUNKSIZE);
|
|
|
296 memcpy(rh->ec_chunk1[x],rh->chans[x].writechunk,ZT_CHUNKSIZE);
|
|
|
297 }
|
|
|
298
|
|
|
299 zt_receive(&rh->span);
|
|
|
300 }
|
|
|
301 ---------------------------------------------------------------------
|
|
|
302
|
|
|
303 **Solution 2:** Bob Conklin from Rhino has suggested this fix that
|
|
|
304 they have in SVN trunk:
|
|
|
305
|
|
|
306 ---------------------------------------------------------------------
|
|
|
307 static void r1t1_receiveprep(struct r1t1 *rh, int nextbuf)
|
|
|
308 {
|
|
|
309 if (!rh->dsp_up)
|
|
|
310 zt_ec_span(&rh->span);
|
|
|
311 zt_receive(&rh->span);
|
|
|
312 }
|
|
|
313 ---------------------------------------------------------------------
|
|
|
314
|
|
|
315 This version automatically falls back to software echo cancellation if
|
|
|
316 the DSP-based hardware echo cancellation is not present. Thanks
|
|
|
317 Bob!
|
|
|
318
|
|
|
319 [[status]]
|
|
|
320 Status
|
|
|
321 ------
|
|
|
322
|
|
|
323 Current status is considered *Stable*. As of 3 January 2008:
|
|
|
324
|
|
|
325 + Thousands of people (estimated) using Oslec on x86 platforms now.
|
|
|
326 Oslec is stable and works well. Oslec has been included in many
|
|
|
327 distributions, including Debian, Gentoo, Trixbox, Elastix, and
|
|
|
328 Callweaver.
|
|
|
329
|
|
|
330 + Oslec works well on many different types of hardware, from single
|
|
|
331 port $10 X100P cards to dual E1s.
|
|
|
332
|
|
|
333 + Oslec has been partially optimised for the embedded Blackfin
|
|
|
334 DSP/RISC CPU and now runs in real time on the link:ip04.html[IP04 Open
|
|
|
335 Hardware IP-PBX]. Running on several hundred IP04s now.
|
|
|
336
|
|
|
337 + Real time runs well on FXO/FXS lines where the other Zaptel echo
|
|
|
338 cancellers (including the latest MG2) struggle. No tweaks like
|
|
|
339 levels, fxotrain, or even opermode required so far. Hardware used for
|
|
|
340 tests was a TDM400 and X100P card.
|
|
|
341
|
|
|
342 + The simulation code passes many but not all of the G168 tests.
|
|
|
343 Some problems with high level signals near 0dBm0 and some of the G168
|
|
|
344 echo models. Passes maybe 90% of the tests attempted so far, however
|
|
|
345 not all of the tests and/or range of test conditions have been
|
|
|
346 attempted yet. The fails are close calls (like a few dB off), not
|
|
|
347 complete breakdowns of the algorithm.
|
|
|
348
|
|
|
349 + Oslec has been specifically developed to work with low cost line
|
|
|
350 interface hardware like the X100P. The X100P has problems with 60Hz
|
|
|
351 hum and DC offset in the rx signal, which interfere with the echo
|
|
|
352 canceller algorithm. However with the addition of some high pass
|
|
|
353 filters good results with X100P cards have been obtained using Oslec.
|
|
|
354 Line interfaces based on chips such as those from Silicon Labs
|
|
|
355 (e.g. the TDM400) have no DC offset or hum which is one reason they
|
|
|
356 tend to perform better.
|
|
|
357
|
|
|
358 + Oslec has been optimised to deal with specific problems encountered
|
|
|
359 with soft phone clients. Due to the high quality microphones and
|
|
|
360 sound blaster hardware used on PCs, more low frequency energy is
|
|
|
361 present compared to normal telephone signals. When this low frequency
|
|
|
362 energy is sent to a hybrid it can force the hybrid into a non linear
|
|
|
363 mode. Echo cancellers assume a linear hybrid so any non-linearities
|
|
|
364 make the echo canceller fall over. The solution is to high pass
|
|
|
365 filter energy sent to the hybrid to ensure excessive low frequency
|
|
|
366 energy is removed. This means inserting a HP filter in the tx path.
|
|
|
367
|
|
|
368 [[directories]]
|
|
|
369 Directories
|
|
|
370 -----------
|
|
|
371
|
|
|
372 The http://svn.astfin.org/software/oslec/trunk/[Oslec SVN repository]
|
|
|
373 combines several projects (mainly spandsp) used to develop, test, and
|
|
|
374 run oslec.
|
|
|
375
|
|
|
376 spandsp : A subset of spandsp to support testing and development of
|
|
|
377 Oslec. The echo canceller and G168 test suite source code
|
|
|
378 live in here.
|
|
|
379
|
|
|
380 user....: User mode apps, e.g. sampling of echo signals and a unit
|
|
|
381 test for measuring the execution speed of Oslec.
|
|
|
382
|
|
|
383 kernel..: Builds Oslec into a kernel module. Also contains patches for
|
|
|
384 Zaptel to allow the use of Oslec.
|
|
|
385
|
|
|
386 [[diagnostics]]
|
|
|
387 Run Time Information
|
|
|
388 --------------------
|
|
|
389
|
|
|
390 Set up a call that uses an analog port, then check out /proc/oslec for
|
|
|
391 real-time stats:
|
|
|
392
|
|
|
393 [root@homework kernel]# cat /proc/oslec/info
|
|
|
394 channels....: 1
|
|
|
395 mode........: [13] |ADAPTION|NLP|CNG|
|
|
|
396 Ltx.........: 0
|
|
|
397 Lrx.........: 211
|
|
|
398 Lclean......: 211
|
|
|
399 Lclean_bg...: 211
|
|
|
400 shift.......: 0
|
|
|
401 Double Talk.: 1
|
|
|
402 MIPs (last)....: 1
|
|
|
403 MIPs (worst)...: 7
|
|
|
404 MIPs (avergage): 1
|
|
|
405
|
|
|
406 You can turn the various mode switches off and on, for example:
|
|
|
407
|
|
|
408 echo 9 > /proc/oslec/mode
|
|
|
409
|
|
|
410 turns off comfort noise, but keeps ADAPT and NLP on. The mode
|
|
|
411 switches are listed in the
|
|
|
412 http://svn.astfin.org/software/oslec/trunk/spandsp-0.0.3/src/spandsp/echo.h[echo.h]
|
|
|
413 header file (#define ECHO_CAN_USE_*). The /proc interface only
|
|
|
414 monitors the first Zaptel call you bring up, see
|
|
|
415 http://svn.astfin.org/software/oslec/trunk/kernel/oslec_wrap.c[oslec_wrap.c]
|
|
|
416 for more information.
|
|
|
417
|
|
|
418 There is a GUI for run-time control of Oslec, called the Oslec Control
|
|
|
419 Panel. For example you can Enable and Disable the echo canceller in
|
|
|
420 real time, as you are speaking. If you Reset the echo canceller you
|
|
|
421 can hear how long it takes to converge again.
|
|
|
422
|
|
|
423 NOTE: The command line tool *dialog* must be installed on your system
|
|
|
424 to use the Oslec Control Panel.
|
|
|
425
|
|
|
426 $ cd kernel
|
|
|
427 $ ./oslec-ctrl-panel.sh
|
|
|
428
|
|
|
429 image::/images/echo/oslec_control_panel.png[Oslec Control Panel]
|
|
|
430
|
|
|
431 [[speed]]
|
|
|
432 Execution Speed and Optimisation
|
|
|
433 --------------------------------
|
|
|
434
|
|
|
435 In the user directory there is a
|
|
|
436 http://svn.astfin.org/software/oslec/trunk/user/speedtest.c[speedtest.c]
|
|
|
437 program that measures how many MIPs the echo canceller consumes and
|
|
|
438 estimates how many cancellers can run in real time. The current x86
|
|
|
439 code is vanilla C and could be greatly improved with MMX or SSE
|
|
|
440 optimisations, some of which are already coded in spandsp. There is
|
|
|
441 also much that can be done to improve execution speed on the Blackfin.
|
|
|
442
|
|
|
443 Steve's spandsp code includes some support for MMX and SSE2
|
|
|
444 optimisation. Define USE_MMX or USE_SSE2 in the Makefile to use this
|
|
|
445 option. At this stage only the fir.h filter code is optimised, oslec
|
|
|
446 would run much faster if lms_adapt_bg() in echo.c was also ported to
|
|
|
447 MMX/SSE2. With USE_MMX defined, the speedtest.c program dropped from
|
|
|
448 33 MIPs to 20 MIPs per channel for 256 taps/32ms. If using MMX in the
|
|
|
449 kernel (e.g. with zaptel) make sure you compile zaptel with
|
|
|
450 CONFIG_ZAPTEL_MMX to ensure the FPU state is saved in the right
|
|
|
451 places.
|
|
|
452
|
|
|
453 To enable MMX/SSE modify this line of oslec/kernel/Makefile:
|
|
|
454
|
|
|
455 ----------------------------------------------
|
|
|
456
|
|
|
457 all::
|
|
|
458 $(MAKE) -C $(KDIR) EXTRA_CFLAGS='$(KINCLUDE) \
|
|
|
459 -DUSE_MMX -DUSE_SSE2 -DEXPORT_SYMTAB -O6' \
|
|
|
460 SUBDIRS=$(PWD) modules
|
|
|
461
|
|
|
462 ----------------------------------------------
|
|
|
463
|
|
|
464 Thanks Nic Bellamy for help in testing MMX and Ming-Ching Tiew for
|
|
|
465 providing the correct command line for compiling with MMX/SSE.
|
|
|
466
|
|
|
467 Some notes on further optimisation:
|
|
|
468
|
|
|
469 1/ How do lots of MIPs in ISR affect total system performance? For
|
|
|
470 example if we are using 25% of MIP in ISR does * still run OK? Find the
|
|
|
471 CPU load knee different for user versus kernel mode cycles
|
|
|
472 consumption.
|
|
|
473
|
|
|
474 2/ Estimated Oslec MIPs are 5(N)(Fs), N is the filter size (number of
|
|
|
475 taps), Fs=8000 is the sampling rate. Factor of 5 is comprised of 1
|
|
|
476 for each FIR filter (forgeround and background), 2 for LMS, 1 for overhead.
|
|
|
477 This suggests that when optimised around 10 MIPs for a 32ms tail.
|
|
|
478
|
|
|
479 3/ There are some more notes on optimisation for the Blackfin in
|
|
|
480 http://svn.astfin.org/software/oslec/trunk/spandsp-0.0.3/src/echo.c[echo.c].
|
|
|
481
|
|
|
482 [[credits]]
|
|
|
483 Background and Credits
|
|
|
484 ----------------------
|
|
|
485
|
|
|
486 Oslec started life as a prototype echo canceller and G168 test
|
|
|
487 framework from Steve Underwood's http://www.soft-switch.org/[Spandsp]
|
|
|
488 library. Steve wrote much of the DSP code used in Asterisk, and the
|
|
|
489 Zaptel echo cancellation code is heavily based on his work.
|
|
|
490
|
|
|
491 Using the Spandsp G168 test framework, a high performance echo
|
|
|
492 canceller has been developed and carefully tested. Working together
|
|
|
493 with alpha testers the performance was brought to a beta state in June
|
|
|
494 2005. <<links,More>>. Since then many thousands of people have
|
|
|
495 installed Oslec.
|
|
|
496
|
|
|
497 Thanks to Steve Underwood, Jean-Marc Valin, and Ramakrishnan
|
|
|
498 Muthukrishnan for their suggestions and email discussions. Thanks
|
|
|
499 also to those people who collected echo samples for me such as Mark,
|
|
|
500 Pawel, and Pavel. Thanks Nic Bellamy for help with testing,
|
|
|
501 explanation of long path issues, and MMX support. Thanks Tzafrir for
|
|
|
502 testing and patches/enhancements and Dmitry for help with
|
|
|
503 multithreaded and locking issues.
|
|
|
504
|
|
|
505 Thanks Vieri for finding a circuit with a tail > 32ms, and Patrick for
|
|
|
506 submitting a Zaptel 1.2.18 patch. Thanks Dave Fullerton and Carlton
|
|
|
507 O'Riley for Zaptel 1.4.3/1.4.4 patches. Thanks Bill Salibrici for
|
|
|
508 finding a memory leak. Thanks Michael E. Kromer for submitting a
|
|
|
509 Zaptel 1.4.7.1 patch, and Tzafrir for the Zaptel 1.4.8 patch. Thanks
|
|
|
510 Russ Price for the Zaptel 1.2.24 patch. Thanks Chris Notley for the
|
|
|
511 1.4.11 patch.
|
|
|
512
|
|
|
513 Thanks to Peter Schlaile for porting Oslec to mISDN. Thanks also to
|
|
|
514 Kristijan Vrban for sending me some ISDN hardware so I can help with
|
|
|
515 the ISDN/Oslec testing!
|
|
|
516
|
|
|
517 Thanks to Tzafrir and Faidon for helping debug the muted audio
|
|
|
518 problems on 64 bit systems.
|
|
|
519
|
|
|
520 Thanks Rudolf E. Steiner for testing Oslec on SMP and multiple E1
|
|
|
521 systems.
|
|
|
522
|
|
|
523 [[sample]]
|
|
|
524 Sampling Zaptel Echo
|
|
|
525 --------------------
|
|
|
526
|
|
|
527 Introduction
|
|
|
528 ~~~~~~~~~~~~
|
|
|
529
|
|
|
530 I have developed a system for sampling echo on running Zaptel/Asterisk
|
|
|
531 systems. If you do experience any echo with Oslec, please use this
|
|
|
532 system to sample the echo, then email the sample files to me.
|
|
|
533
|
|
|
534 See the http://www.rowetel.com/blog/?p=18[Part 1 blog post] for more
|
|
|
535 information on sampling echo.
|
|
|
536
|
|
|
537 The nice thing about is that it doesn't interfere with your running
|
|
|
538 Asterisk system. If you hear echo at any time you can fire up a
|
|
|
539 console and run "sample" to capture real-time data from the Zaptel
|
|
|
540 port.
|
|
|
541
|
|
|
542 There are other ways to capture audio from a running Asterisk system,
|
|
|
543 for example using the built in Asterisk play and record applications.
|
|
|
544 However this system is designed to capture *exactly* the signals being
|
|
|
545 fed to & from the echo canceller - preserving the exact timing of the
|
|
|
546 signals and with no intermediate buffering. This is very important
|
|
|
547 for echo cancellation work - the algorithms depend on an exact timing
|
|
|
548 relationship between the transmit and receive signals.
|
|
|
549
|
|
|
550 The system simultaneously captures the receive signal both before
|
|
|
551 *and* after the echo canceller - something that is difficult to do
|
|
|
552 with the built in Asterisk functions due to the location of the echo
|
|
|
553 canceller deep in the Zaptel driver.
|
|
|
554
|
|
|
555 If you would like to help further develop Oslec, please
|
|
|
556 mailto:david_at_rowetel_dot_com[send me] your echo samples! I would
|
|
|
557 welcome any samples of your echo signals, for example where the echo
|
|
|
558 canceller isn't working well, and also cases where it does work
|
|
|
559 well. By comparing the two cases we can learn a lot about the
|
|
|
560 strengths and weaknesses of the algorithm.
|
|
|
561
|
|
|
562 Installing
|
|
|
563 ~~~~~~~~~~
|
|
|
564
|
|
|
565 1/ Install Oslec ("HowTo - Run OSLEC with Asterisk/Zaptel" section above).
|
|
|
566
|
|
|
567 2/ If you would like to use zaptap without oslec change the selected
|
|
|
568 echo canceller in zconfig.c and rebuild/install zaptel.
|
|
|
569
|
|
|
570 3/ Compile sample.c:
|
|
|
571
|
|
|
572 # cd user
|
|
|
573 # make
|
|
|
574
|
|
|
575 4/ Create a device node. I used major number 33 as it was
|
|
|
576 free on my PC:
|
|
|
577
|
|
|
578 # mknod -m 666 /dev/sample c 33 0
|
|
|
579
|
|
|
580 If 33 is not free choose a free major number and change the
|
|
|
581 #define SAMPLE_MAJOR in zaptel.c and recompile.
|
|
|
582
|
|
|
583 Note it is important to "insmod zaptel.o" from inside the zaptel
|
|
|
584 directory and don't forget the ".o" (.ko for Linux 2.6). Otherwise
|
|
|
585 insmod will use the previously installed version of zaptel and
|
|
|
586 "sample" won't work.
|
|
|
587
|
|
|
588 5/ Make the call to the Zaptel port you wish to sample. Run
|
|
|
589 sample while talking:
|
|
|
590
|
|
|
591 # ./sample test 1 5
|
|
|
592
|
|
|
593 this will create test_tx.raw, test_rx.raw & test_ec.raw. There
|
|
|
594 will be a few messages on the console as the driver does it's
|
|
|
595 thing, you can check these with dmesg.
|
|
|
596
|
|
|
597 6/ You can check your samples by playing them back through your
|
|
|
598 sound card, for example:
|
|
|
599
|
|
|
600 # play -f s -r 8000 -s w test_ec.raw
|
|
|
601
|
|
|
602 7/ To convert the raw files to wave files (this is more convenient for playing
|
|
|
603 and processing with the Oslec simulation):
|
|
|
604
|
|
|
605 # sox -t raw -r 8000 -s -w -c 1 test_ec.raw test_ec.wav
|
|
|
606 # play test_ec.wav
|
|
|
607
|
|
|
608 8/ There is an Octave script pl.m to help plot the samples:
|
|
|
609
|
|
|
610 # cp /your/test/samples ~/oslec/user
|
|
|
611 # cd ~/oslec/user
|
|
|
612 # octave
|
|
|
613 octave:1> pl("test")
|
|
|
614
|
|
|
615 Configuration
|
|
|
616 ~~~~~~~~~~~~~
|
|
|
617
|
|
|
618 + Zaptel/Asterisk 1.2.13, 1.4.0 or 1.4.1. However other versions should work
|
|
|
619 OK.
|
|
|
620 + Linux 2.4 or 2.6 Kernel
|
|
|
621 + Digium TDM400 hardware
|
|
|
622 + My zapata.conf for the FXO port under test is something like:
|
|
|
623
|
|
|
624 ---------------------------------------------------------------------
|
|
|
625 signalling=fxs_ks
|
|
|
626 echocancel=yes
|
|
|
627 ;echocancelwhenbridged=yes
|
|
|
628 ;echotraining=400
|
|
|
629 group=2;
|
|
|
630 conext=incoming
|
|
|
631 channel =>4
|
|
|
632 ---------------------------------------------------------------------
|
|
|
633
|
|
|
634 Note echo training is switched off, although this didn't make much
|
|
|
635 difference to the FXO Port when using the built-in Zaptel echo
|
|
|
636 cancellers. (helped the the FXS port though). Oslec does not support
|
|
|
637 echo training.
|
|
|
638
|
|
|
639 [[support]]
|
|
|
640 Support
|
|
|
641 -------
|
|
|
642
|
|
|
643 https://lists.sourceforge.net/lists/listinfo/freetel-oslec[Oslec
|
|
|
644 mailing list].
|
|
|
645
|
|
|
646 [[further]]
|
|
|
647 Further Work
|
|
|
648 ------------
|
|
|
649
|
|
|
650 Here are some ideas for further work.
|
|
|
651
|
|
|
652 1/ Make info screen on control panel update each second automatically.
|
|
|
653
|
|
|
654 2/ When e/c is reset during a call it converges faster than at start of
|
|
|
655 calls. This suggests e/c is wandering off into dumb states before
|
|
|
656 the call is connected. Perhaps disabling adaption until the call
|
|
|
657 is connected would help with faster convergence
|
|
|
658
|
|
|
659 3/ Preserving the state of the e/c between calls is also a very good
|
|
|
660 idea, current zaptel design destroys the e/c at the end of a call.
|
|
|
661
|
|
|
662 4/ Switch on and test SSE and MMX filter code for x86. Code LMS
|
|
|
663 update for x86 in SSE or MMX.
|
|
|
664
|
|
|
665 5/ Add a feature to /proc/oslec to extract the current estimated
|
|
|
666 impulse response. Use gnuplot or Octave to plot it in real time.
|
|
|
667
|
|
|
668 6/ Attempt to speed convergence.
|
|
|
669
|
|
|
670 7/ Set up test scripts for greater coverage of G168 tests.
|
|
|
671
|
|
|
672 8/ If necessary develop a sparse approach to handle 128ms tails.
|
|
|
673
|
|
|
674 [[Simulation]]
|
|
|
675 HowTo - OSLEC G168 Simulation
|
|
|
676 -----------------------------
|
|
|
677
|
|
|
678 The simulation form of Oslec is useful for Oslec development. It is
|
|
|
679 much easier to develop using a non-real time, user mode program than
|
|
|
680 saying 'Hello,...1,...2,....3" down a telephone line. The simulation
|
|
|
681 dumps internal states while it is running, which can be plotted and
|
|
|
682 analysed using Octave.
|
|
|
683
|
|
|
684 1. Read spandsp-0.0.3/README and make sure you have installed the
|
|
|
685 dependencies like libtiff-devel, libaudiofile-devel, fftw-devel.
|
|
|
686
|
|
|
687 2. Make, build, run a test:
|
|
|
688
|
|
|
689 $ cd spandsp-0.0.3;
|
|
|
690 $ ./configure --enable-tests
|
|
|
691 $ make
|
|
|
692 $ cd tests
|
|
|
693 $ ./echo_tests 2a
|
|
|
694
|
|
|
695 3. You can change the echo path model with -m [1..7] (default 1)
|
|
|
696
|
|
|
697 4. You can change the ERL with -erl [0..whatever] (default 10.0)
|
|
|
698
|
|
|
699 5. You can plot the internal states using Octave:
|
|
|
700
|
|
|
701 $ cd spandsp-0.0.3/echo_tests
|
|
|
702 $ octave
|
|
|
703 octave:1> echo_dump
|
|
|
704
|
|
|
705 The st= and en= statements at the top of echo_dump.m control which
|
|
|
706 part of the waveform is plotted.
|
|
|
707
|
|
|
708 6. There are some sequences of tests set up in the script files:
|
|
|
709
|
|
|
710 $ ./g168_quick.sh (useful for a quick sanity test)
|
|
|
711 $ ./g168_tests.sh (more comprehensive set of tests)
|
|
|
712
|
|
|
713 7. For more options:
|
|
|
714
|
|
|
715 $ ./echo_tests -h
|
|
|
716
|
|
|
717 8. Counting passes (say if u want to see how many tests pass after making
|
|
|
718 a change):
|
|
|
719
|
|
|
720 $ ./g168_tests.sh > lms16bit.txt
|
|
|
721 $ cat lms16bit.txt | grep PASS | wc -l
|
|
|
722
|
|
|
723 [[links]]
|
|
|
724 Further Reading
|
|
|
725 ---------------
|
|
|
726
|
|
|
727 + http://svn.astfin.org/software/oslec/trunk/spandsp-0.0.3/src/echo.c[echo.c]
|
|
|
728 is the heart of the Oslec echo canceller.
|
|
|
729
|
|
|
730 There are several blog posts documenting the development of Oslec:
|
|
|
731
|
|
|
732 + http://www.rowetel.com/blog/?p=18[Part 1 - Introduction] discusses the myth
|
|
|
733 of hardware echo cancellation and the concept of an echo canceller
|
|
|
734 developed using echo samples collected by a community.
|
|
|
735
|
|
|
736 + http://www.rowetel.com/blog/?p=21[Part 2 - How Echo Cancellers Work]
|
|
|
737 is an easy to read introduction to echo cancellation for C programmers.
|
|
|
738
|
|
|
739 + http://www.rowetel.com/blog/?p=22[Part 3 - Two Prototypes] discusses
|
|
|
740 two algorithms that were developed as candidates for Oslec. This is a
|
|
|
741 fairly "hard core" DSP post - some familiarity with echo cancellers is
|
|
|
742 assumed.
|
|
|
743
|
|
|
744 + http://www.rowetel.com/blog/?p=23[Part 4 - First Phone calls] talks
|
|
|
745 about the very first real-world phone calls made using Oslec.
|
|
|
746
|
|
|
747 + http://www.rowetel.com/blog/?p=33[Part 5 - Ready for Beta Testing].
|
|
|
748 Walks through some of the alpha testing bugs and how they were fixed,
|
|
|
749 discusses open development methods and the need for 128ms tails.
|
|
|
750
|
|
|
751 Some useful links:
|
|
|
752
|
|
|
753 + Ochiai, Areseki, and Ogihara, "Echo Canceller with Two Echo Path
|
|
|
754 Models", IEEE Transactions on communications, COM-25, No. 6, June 1977.
|
|
|
755 http://www.rowetel.com/images/echo/dual_path_paper.pdf[download].
|
|
|
756
|
|
|
757 + The classic, very useful paper that tells you how to actually build
|
|
|
758 a real world echo canceller: Messerschmitt, Hedberg, Cole, Haoui,
|
|
|
759 Winship, "Digital Voice Echo Canceller with a
|
|
|
760 TMS320020. http://www.rowetel.com/images/echo/spra129.pdf[download].
|
|
|
761
|
|
|
762 + A nice
|
|
|
763 http://en.wikipedia.org/wiki/Least_mean_squares_filter[introduction to
|
|
|
764 LMS filters].
|
|
|
765
|
|
|
766 + Good introduction to http://www.cisco.com/univercd/cc/td/doc/cisintwk/intsolns/voipsol/ea_isd.htm[echo on VOIP calls] from Cisco.
|
|
|
767
|
