]>
Commit | Line | Data |
---|---|---|
10602db8 DR |
1 | /* |
2 | * SpanDSP - a series of DSP components for telephony | |
3 | * | |
4 | * echo.c - A line echo canceller. This code is being developed | |
5 | * against and partially complies with G168. | |
6 | * | |
7 | * Written by Steve Underwood <steveu@coppice.org> | |
8 | * and David Rowe <david_at_rowetel_dot_com> | |
9 | * | |
10 | * Copyright (C) 2001, 2003 Steve Underwood, 2007 David Rowe | |
11 | * | |
12 | * Based on a bit from here, a bit from there, eye of toad, ear of | |
13 | * bat, 15 years of failed attempts by David and a few fried brain | |
14 | * cells. | |
15 | * | |
16 | * All rights reserved. | |
17 | * | |
18 | * This program is free software; you can redistribute it and/or modify | |
19 | * it under the terms of the GNU General Public License version 2, as | |
20 | * published by the Free Software Foundation. | |
21 | * | |
22 | * This program is distributed in the hope that it will be useful, | |
23 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
24 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
25 | * GNU General Public License for more details. | |
26 | * | |
27 | * You should have received a copy of the GNU General Public License | |
28 | * along with this program; if not, write to the Free Software | |
29 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
10602db8 DR |
30 | */ |
31 | ||
32 | /*! \file */ | |
33 | ||
34 | /* Implementation Notes | |
35 | David Rowe | |
36 | April 2007 | |
37 | ||
38 | This code started life as Steve's NLMS algorithm with a tap | |
39 | rotation algorithm to handle divergence during double talk. I | |
40 | added a Geigel Double Talk Detector (DTD) [2] and performed some | |
41 | G168 tests. However I had trouble meeting the G168 requirements, | |
42 | especially for double talk - there were always cases where my DTD | |
43 | failed, for example where near end speech was under the 6dB | |
44 | threshold required for declaring double talk. | |
45 | ||
46 | So I tried a two path algorithm [1], which has so far given better | |
47 | results. The original tap rotation/Geigel algorithm is available | |
48 | in SVN http://svn.rowetel.com/software/oslec/tags/before_16bit. | |
49 | It's probably possible to make it work if some one wants to put some | |
50 | serious work into it. | |
51 | ||
52 | At present no special treatment is provided for tones, which | |
53 | generally cause NLMS algorithms to diverge. Initial runs of a | |
54 | subset of the G168 tests for tones (e.g ./echo_test 6) show the | |
55 | current algorithm is passing OK, which is kind of surprising. The | |
56 | full set of tests needs to be performed to confirm this result. | |
57 | ||
58 | One other interesting change is that I have managed to get the NLMS | |
59 | code to work with 16 bit coefficients, rather than the original 32 | |
60 | bit coefficents. This reduces the MIPs and storage required. | |
61 | I evaulated the 16 bit port using g168_tests.sh and listening tests | |
62 | on 4 real-world samples. | |
63 | ||
64 | I also attempted the implementation of a block based NLMS update | |
65 | [2] but although this passes g168_tests.sh it didn't converge well | |
66 | on the real-world samples. I have no idea why, perhaps a scaling | |
67 | problem. The block based code is also available in SVN | |
68 | http://svn.rowetel.com/software/oslec/tags/before_16bit. If this | |
69 | code can be debugged, it will lead to further reduction in MIPS, as | |
70 | the block update code maps nicely onto DSP instruction sets (it's a | |
71 | dot product) compared to the current sample-by-sample update. | |
72 | ||
73 | Steve also has some nice notes on echo cancellers in echo.h | |
74 | ||
10602db8 DR |
75 | References: |
76 | ||
77 | [1] Ochiai, Areseki, and Ogihara, "Echo Canceller with Two Echo | |
78 | Path Models", IEEE Transactions on communications, COM-25, | |
79 | No. 6, June | |
80 | 1977. | |
81 | http://www.rowetel.com/images/echo/dual_path_paper.pdf | |
82 | ||
83 | [2] The classic, very useful paper that tells you how to | |
84 | actually build a real world echo canceller: | |
49bb9e6d GKH |
85 | Messerschmitt, Hedberg, Cole, Haoui, Winship, "Digital Voice |
86 | Echo Canceller with a TMS320020, | |
87 | http://www.rowetel.com/images/echo/spra129.pdf | |
10602db8 DR |
88 | |
89 | [3] I have written a series of blog posts on this work, here is | |
90 | Part 1: http://www.rowetel.com/blog/?p=18 | |
91 | ||
92 | [4] The source code http://svn.rowetel.com/software/oslec/ | |
93 | ||
94 | [5] A nice reference on LMS filters: | |
49bb9e6d | 95 | http://en.wikipedia.org/wiki/Least_mean_squares_filter |
10602db8 DR |
96 | |
97 | Credits: | |
98 | ||
99 | Thanks to Steve Underwood, Jean-Marc Valin, and Ramakrishnan | |
100 | Muthukrishnan for their suggestions and email discussions. Thanks | |
101 | also to those people who collected echo samples for me such as | |
102 | Mark, Pawel, and Pavel. | |
103 | */ | |
104 | ||
49bb9e6d | 105 | #include <linux/kernel.h> |
10602db8 | 106 | #include <linux/module.h> |
10602db8 | 107 | #include <linux/slab.h> |
10602db8 | 108 | |
10602db8 DR |
109 | #include "echo.h" |
110 | ||
49bb9e6d GKH |
111 | #define MIN_TX_POWER_FOR_ADAPTION 64 |
112 | #define MIN_RX_POWER_FOR_ADAPTION 64 | |
113 | #define DTD_HANGOVER 600 /* 600 samples, or 75ms */ | |
114 | #define DC_LOG2BETA 3 /* log2() of DC filter Beta */ | |
10602db8 | 115 | |
10602db8 DR |
116 | |
117 | /* adapting coeffs using the traditional stochastic descent (N)LMS algorithm */ | |
118 | ||
f55ccbf6 | 119 | #ifdef __bfin__ |
dc57a3ea | 120 | static inline void lms_adapt_bg(struct oslec_state *ec, int clean, |
4460a860 | 121 | int shift) |
10602db8 | 122 | { |
4460a860 M |
123 | int i, j; |
124 | int offset1; | |
125 | int offset2; | |
126 | int factor; | |
127 | int exp; | |
128 | int16_t *phist; | |
129 | int n; | |
130 | ||
131 | if (shift > 0) | |
132 | factor = clean << shift; | |
133 | else | |
134 | factor = clean >> -shift; | |
135 | ||
136 | /* Update the FIR taps */ | |
137 | ||
138 | offset2 = ec->curr_pos; | |
139 | offset1 = ec->taps - offset2; | |
140 | phist = &ec->fir_state_bg.history[offset2]; | |
141 | ||
142 | /* st: and en: help us locate the assembler in echo.s */ | |
143 | ||
dc57a3ea | 144 | /* asm("st:"); */ |
4460a860 M |
145 | n = ec->taps; |
146 | for (i = 0, j = offset2; i < n; i++, j++) { | |
147 | exp = *phist++ * factor; | |
148 | ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15); | |
149 | } | |
dc57a3ea | 150 | /* asm("en:"); */ |
4460a860 M |
151 | |
152 | /* Note the asm for the inner loop above generated by Blackfin gcc | |
153 | 4.1.1 is pretty good (note even parallel instructions used): | |
154 | ||
155 | R0 = W [P0++] (X); | |
156 | R0 *= R2; | |
157 | R0 = R0 + R3 (NS) || | |
158 | R1 = W [P1] (X) || | |
159 | nop; | |
160 | R0 >>>= 15; | |
161 | R0 = R0 + R1; | |
162 | W [P1++] = R0; | |
163 | ||
164 | A block based update algorithm would be much faster but the | |
165 | above can't be improved on much. Every instruction saved in | |
166 | the loop above is 2 MIPs/ch! The for loop above is where the | |
167 | Blackfin spends most of it's time - about 17 MIPs/ch measured | |
168 | with speedtest.c with 256 taps (32ms). Write-back and | |
169 | Write-through cache gave about the same performance. | |
170 | */ | |
10602db8 DR |
171 | } |
172 | ||
173 | /* | |
174 | IDEAS for further optimisation of lms_adapt_bg(): | |
175 | ||
176 | 1/ The rounding is quite costly. Could we keep as 32 bit coeffs | |
177 | then make filter pluck the MS 16-bits of the coeffs when filtering? | |
178 | However this would lower potential optimisation of filter, as I | |
179 | think the dual-MAC architecture requires packed 16 bit coeffs. | |
180 | ||
181 | 2/ Block based update would be more efficient, as per comments above, | |
182 | could use dual MAC architecture. | |
183 | ||
184 | 3/ Look for same sample Blackfin LMS code, see if we can get dual-MAC | |
185 | packing. | |
186 | ||
187 | 4/ Execute the whole e/c in a block of say 20ms rather than sample | |
188 | by sample. Processing a few samples every ms is inefficient. | |
189 | */ | |
190 | ||
191 | #else | |
dc57a3ea | 192 | static inline void lms_adapt_bg(struct oslec_state *ec, int clean, |
4460a860 | 193 | int shift) |
10602db8 | 194 | { |
4460a860 M |
195 | int i; |
196 | ||
197 | int offset1; | |
198 | int offset2; | |
199 | int factor; | |
200 | int exp; | |
201 | ||
202 | if (shift > 0) | |
203 | factor = clean << shift; | |
204 | else | |
205 | factor = clean >> -shift; | |
206 | ||
207 | /* Update the FIR taps */ | |
208 | ||
209 | offset2 = ec->curr_pos; | |
210 | offset1 = ec->taps - offset2; | |
211 | ||
212 | for (i = ec->taps - 1; i >= offset1; i--) { | |
213 | exp = (ec->fir_state_bg.history[i - offset1] * factor); | |
214 | ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15); | |
215 | } | |
216 | for (; i >= 0; i--) { | |
217 | exp = (ec->fir_state_bg.history[i + offset2] * factor); | |
218 | ec->fir_taps16[1][i] += (int16_t) ((exp + (1 << 14)) >> 15); | |
219 | } | |
10602db8 DR |
220 | } |
221 | #endif | |
222 | ||
56791f0a | 223 | static inline int top_bit(unsigned int bits) |
196e76e8 DR |
224 | { |
225 | if (bits == 0) | |
56791f0a GKH |
226 | return -1; |
227 | else | |
228 | return (int)fls((int32_t)bits)-1; | |
196e76e8 DR |
229 | } |
230 | ||
9d8f2d5d | 231 | struct oslec_state *oslec_create(int len, int adaption_mode) |
10602db8 | 232 | { |
4460a860 M |
233 | struct oslec_state *ec; |
234 | int i; | |
235 | ||
236 | ec = kzalloc(sizeof(*ec), GFP_KERNEL); | |
237 | if (!ec) | |
238 | return NULL; | |
239 | ||
240 | ec->taps = len; | |
241 | ec->log2taps = top_bit(len); | |
242 | ec->curr_pos = ec->taps - 1; | |
243 | ||
244 | for (i = 0; i < 2; i++) { | |
245 | ec->fir_taps16[i] = | |
246 | kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL); | |
247 | if (!ec->fir_taps16[i]) | |
248 | goto error_oom; | |
249 | } | |
250 | ||
251 | fir16_create(&ec->fir_state, ec->fir_taps16[0], ec->taps); | |
252 | fir16_create(&ec->fir_state_bg, ec->fir_taps16[1], ec->taps); | |
253 | ||
dc57a3ea | 254 | for (i = 0; i < 5; i++) |
4460a860 | 255 | ec->xvtx[i] = ec->yvtx[i] = ec->xvrx[i] = ec->yvrx[i] = 0; |
4460a860 M |
256 | |
257 | ec->cng_level = 1000; | |
258 | oslec_adaption_mode(ec, adaption_mode); | |
259 | ||
260 | ec->snapshot = kcalloc(ec->taps, sizeof(int16_t), GFP_KERNEL); | |
261 | if (!ec->snapshot) | |
262 | goto error_oom; | |
263 | ||
264 | ec->cond_met = 0; | |
265 | ec->Pstates = 0; | |
266 | ec->Ltxacc = ec->Lrxacc = ec->Lcleanacc = ec->Lclean_bgacc = 0; | |
267 | ec->Ltx = ec->Lrx = ec->Lclean = ec->Lclean_bg = 0; | |
268 | ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0; | |
269 | ec->Lbgn = ec->Lbgn_acc = 0; | |
270 | ec->Lbgn_upper = 200; | |
271 | ec->Lbgn_upper_acc = ec->Lbgn_upper << 13; | |
272 | ||
273 | return ec; | |
274 | ||
dc57a3ea | 275 | error_oom: |
4460a860 M |
276 | for (i = 0; i < 2; i++) |
277 | kfree(ec->fir_taps16[i]); | |
278 | ||
279 | kfree(ec); | |
280 | return NULL; | |
10602db8 | 281 | } |
9d8f2d5d | 282 | EXPORT_SYMBOL_GPL(oslec_create); |
10602db8 | 283 | |
9d8f2d5d | 284 | void oslec_free(struct oslec_state *ec) |
10602db8 DR |
285 | { |
286 | int i; | |
287 | ||
288 | fir16_free(&ec->fir_state); | |
289 | fir16_free(&ec->fir_state_bg); | |
4460a860 | 290 | for (i = 0; i < 2; i++) |
10602db8 DR |
291 | kfree(ec->fir_taps16[i]); |
292 | kfree(ec->snapshot); | |
293 | kfree(ec); | |
294 | } | |
9d8f2d5d | 295 | EXPORT_SYMBOL_GPL(oslec_free); |
10602db8 | 296 | |
9d8f2d5d | 297 | void oslec_adaption_mode(struct oslec_state *ec, int adaption_mode) |
10602db8 | 298 | { |
4460a860 | 299 | ec->adaption_mode = adaption_mode; |
10602db8 | 300 | } |
9d8f2d5d | 301 | EXPORT_SYMBOL_GPL(oslec_adaption_mode); |
10602db8 | 302 | |
9d8f2d5d | 303 | void oslec_flush(struct oslec_state *ec) |
10602db8 | 304 | { |
4460a860 | 305 | int i; |
10602db8 | 306 | |
4460a860 M |
307 | ec->Ltxacc = ec->Lrxacc = ec->Lcleanacc = ec->Lclean_bgacc = 0; |
308 | ec->Ltx = ec->Lrx = ec->Lclean = ec->Lclean_bg = 0; | |
309 | ec->tx_1 = ec->tx_2 = ec->rx_1 = ec->rx_2 = 0; | |
10602db8 | 310 | |
4460a860 M |
311 | ec->Lbgn = ec->Lbgn_acc = 0; |
312 | ec->Lbgn_upper = 200; | |
313 | ec->Lbgn_upper_acc = ec->Lbgn_upper << 13; | |
10602db8 | 314 | |
4460a860 | 315 | ec->nonupdate_dwell = 0; |
10602db8 | 316 | |
4460a860 M |
317 | fir16_flush(&ec->fir_state); |
318 | fir16_flush(&ec->fir_state_bg); | |
319 | ec->fir_state.curr_pos = ec->taps - 1; | |
320 | ec->fir_state_bg.curr_pos = ec->taps - 1; | |
321 | for (i = 0; i < 2; i++) | |
322 | memset(ec->fir_taps16[i], 0, ec->taps * sizeof(int16_t)); | |
10602db8 | 323 | |
4460a860 M |
324 | ec->curr_pos = ec->taps - 1; |
325 | ec->Pstates = 0; | |
10602db8 | 326 | } |
9d8f2d5d | 327 | EXPORT_SYMBOL_GPL(oslec_flush); |
10602db8 | 328 | |
4460a860 M |
329 | void oslec_snapshot(struct oslec_state *ec) |
330 | { | |
331 | memcpy(ec->snapshot, ec->fir_taps16[0], ec->taps * sizeof(int16_t)); | |
10602db8 | 332 | } |
9d8f2d5d | 333 | EXPORT_SYMBOL_GPL(oslec_snapshot); |
10602db8 | 334 | |
49bb9e6d | 335 | /* Dual Path Echo Canceller */ |
10602db8 | 336 | |
9d8f2d5d | 337 | int16_t oslec_update(struct oslec_state *ec, int16_t tx, int16_t rx) |
10602db8 | 338 | { |
4460a860 M |
339 | int32_t echo_value; |
340 | int clean_bg; | |
341 | int tmp, tmp1; | |
342 | ||
49bb9e6d GKH |
343 | /* |
344 | * Input scaling was found be required to prevent problems when tx | |
345 | * starts clipping. Another possible way to handle this would be the | |
346 | * filter coefficent scaling. | |
347 | */ | |
4460a860 M |
348 | |
349 | ec->tx = tx; | |
350 | ec->rx = rx; | |
351 | tx >>= 1; | |
352 | rx >>= 1; | |
353 | ||
354 | /* | |
49bb9e6d GKH |
355 | * Filter DC, 3dB point is 160Hz (I think), note 32 bit precision |
356 | * required otherwise values do not track down to 0. Zero at DC, Pole | |
196e76e8 | 357 | * at (1-Beta) on real axis. Some chip sets (like Si labs) don't |
49bb9e6d GKH |
358 | * need this, but something like a $10 X100P card does. Any DC really |
359 | * slows down convergence. | |
360 | * | |
361 | * Note: removes some low frequency from the signal, this reduces the | |
362 | * speech quality when listening to samples through headphones but may | |
363 | * not be obvious through a telephone handset. | |
364 | * | |
365 | * Note that the 3dB frequency in radians is approx Beta, e.g. for Beta | |
366 | * = 2^(-3) = 0.125, 3dB freq is 0.125 rads = 159Hz. | |
4460a860 M |
367 | */ |
368 | ||
369 | if (ec->adaption_mode & ECHO_CAN_USE_RX_HPF) { | |
370 | tmp = rx << 15; | |
196e76e8 | 371 | |
49bb9e6d GKH |
372 | /* |
373 | * Make sure the gain of the HPF is 1.0. This can still | |
374 | * saturate a little under impulse conditions, and it might | |
375 | * roll to 32768 and need clipping on sustained peak level | |
376 | * signals. However, the scale of such clipping is small, and | |
377 | * the error due to any saturation should not markedly affect | |
378 | * the downstream processing. | |
379 | */ | |
4460a860 | 380 | tmp -= (tmp >> 4); |
196e76e8 | 381 | |
4460a860 M |
382 | ec->rx_1 += -(ec->rx_1 >> DC_LOG2BETA) + tmp - ec->rx_2; |
383 | ||
49bb9e6d GKH |
384 | /* |
385 | * hard limit filter to prevent clipping. Note that at this | |
386 | * stage rx should be limited to +/- 16383 due to right shift | |
387 | * above | |
388 | */ | |
4460a860 M |
389 | tmp1 = ec->rx_1 >> 15; |
390 | if (tmp1 > 16383) | |
391 | tmp1 = 16383; | |
392 | if (tmp1 < -16383) | |
393 | tmp1 = -16383; | |
394 | rx = tmp1; | |
395 | ec->rx_2 = tmp; | |
396 | } | |
10602db8 | 397 | |
4460a860 M |
398 | /* Block average of power in the filter states. Used for |
399 | adaption power calculation. */ | |
10602db8 | 400 | |
4460a860 M |
401 | { |
402 | int new, old; | |
403 | ||
404 | /* efficient "out with the old and in with the new" algorithm so | |
405 | we don't have to recalculate over the whole block of | |
406 | samples. */ | |
dc57a3ea | 407 | new = (int)tx * (int)tx; |
4460a860 M |
408 | old = (int)ec->fir_state.history[ec->fir_state.curr_pos] * |
409 | (int)ec->fir_state.history[ec->fir_state.curr_pos]; | |
410 | ec->Pstates += | |
0f51010e | 411 | ((new - old) + (1 << (ec->log2taps-1))) >> ec->log2taps; |
4460a860 M |
412 | if (ec->Pstates < 0) |
413 | ec->Pstates = 0; | |
414 | } | |
10602db8 | 415 | |
4460a860 | 416 | /* Calculate short term average levels using simple single pole IIRs */ |
10602db8 | 417 | |
4460a860 M |
418 | ec->Ltxacc += abs(tx) - ec->Ltx; |
419 | ec->Ltx = (ec->Ltxacc + (1 << 4)) >> 5; | |
420 | ec->Lrxacc += abs(rx) - ec->Lrx; | |
421 | ec->Lrx = (ec->Lrxacc + (1 << 4)) >> 5; | |
10602db8 | 422 | |
49bb9e6d | 423 | /* Foreground filter */ |
10602db8 | 424 | |
4460a860 M |
425 | ec->fir_state.coeffs = ec->fir_taps16[0]; |
426 | echo_value = fir16(&ec->fir_state, tx); | |
427 | ec->clean = rx - echo_value; | |
428 | ec->Lcleanacc += abs(ec->clean) - ec->Lclean; | |
429 | ec->Lclean = (ec->Lcleanacc + (1 << 4)) >> 5; | |
10602db8 | 430 | |
49bb9e6d | 431 | /* Background filter */ |
10602db8 | 432 | |
4460a860 M |
433 | echo_value = fir16(&ec->fir_state_bg, tx); |
434 | clean_bg = rx - echo_value; | |
435 | ec->Lclean_bgacc += abs(clean_bg) - ec->Lclean_bg; | |
436 | ec->Lclean_bg = (ec->Lclean_bgacc + (1 << 4)) >> 5; | |
10602db8 | 437 | |
49bb9e6d | 438 | /* Background Filter adaption */ |
10602db8 | 439 | |
4460a860 M |
440 | /* Almost always adap bg filter, just simple DT and energy |
441 | detection to minimise adaption in cases of strong double talk. | |
442 | However this is not critical for the dual path algorithm. | |
443 | */ | |
444 | ec->factor = 0; | |
445 | ec->shift = 0; | |
446 | if ((ec->nonupdate_dwell == 0)) { | |
447 | int P, logP, shift; | |
448 | ||
449 | /* Determine: | |
450 | ||
451 | f = Beta * clean_bg_rx/P ------ (1) | |
452 | ||
453 | where P is the total power in the filter states. | |
454 | ||
455 | The Boffins have shown that if we obey (1) we converge | |
456 | quickly and avoid instability. | |
457 | ||
458 | The correct factor f must be in Q30, as this is the fixed | |
459 | point format required by the lms_adapt_bg() function, | |
460 | therefore the scaled version of (1) is: | |
461 | ||
462 | (2^30) * f = (2^30) * Beta * clean_bg_rx/P | |
196e76e8 | 463 | factor = (2^30) * Beta * clean_bg_rx/P ----- (2) |
4460a860 M |
464 | |
465 | We have chosen Beta = 0.25 by experiment, so: | |
466 | ||
196e76e8 | 467 | factor = (2^30) * (2^-2) * clean_bg_rx/P |
4460a860 | 468 | |
56791f0a | 469 | (30 - 2 - log2(P)) |
196e76e8 | 470 | factor = clean_bg_rx 2 ----- (3) |
4460a860 M |
471 | |
472 | To avoid a divide we approximate log2(P) as top_bit(P), | |
473 | which returns the position of the highest non-zero bit in | |
474 | P. This approximation introduces an error as large as a | |
475 | factor of 2, but the algorithm seems to handle it OK. | |
476 | ||
477 | Come to think of it a divide may not be a big deal on a | |
478 | modern DSP, so its probably worth checking out the cycles | |
479 | for a divide versus a top_bit() implementation. | |
480 | */ | |
481 | ||
482 | P = MIN_TX_POWER_FOR_ADAPTION + ec->Pstates; | |
483 | logP = top_bit(P) + ec->log2taps; | |
484 | shift = 30 - 2 - logP; | |
485 | ec->shift = shift; | |
486 | ||
487 | lms_adapt_bg(ec, clean_bg, shift); | |
10602db8 | 488 | } |
4460a860 M |
489 | |
490 | /* very simple DTD to make sure we dont try and adapt with strong | |
491 | near end speech */ | |
492 | ||
493 | ec->adapt = 0; | |
494 | if ((ec->Lrx > MIN_RX_POWER_FOR_ADAPTION) && (ec->Lrx > ec->Ltx)) | |
495 | ec->nonupdate_dwell = DTD_HANGOVER; | |
496 | if (ec->nonupdate_dwell) | |
497 | ec->nonupdate_dwell--; | |
498 | ||
49bb9e6d | 499 | /* Transfer logic */ |
4460a860 M |
500 | |
501 | /* These conditions are from the dual path paper [1], I messed with | |
502 | them a bit to improve performance. */ | |
503 | ||
504 | if ((ec->adaption_mode & ECHO_CAN_USE_ADAPTION) && | |
505 | (ec->nonupdate_dwell == 0) && | |
dc57a3ea AB |
506 | /* (ec->Lclean_bg < 0.875*ec->Lclean) */ |
507 | (8 * ec->Lclean_bg < 7 * ec->Lclean) && | |
508 | /* (ec->Lclean_bg < 0.125*ec->Ltx) */ | |
509 | (8 * ec->Lclean_bg < ec->Ltx)) { | |
4460a860 | 510 | if (ec->cond_met == 6) { |
49bb9e6d GKH |
511 | /* |
512 | * BG filter has had better results for 6 consecutive | |
513 | * samples | |
514 | */ | |
4460a860 M |
515 | ec->adapt = 1; |
516 | memcpy(ec->fir_taps16[0], ec->fir_taps16[1], | |
dc57a3ea | 517 | ec->taps * sizeof(int16_t)); |
4460a860 M |
518 | } else |
519 | ec->cond_met++; | |
520 | } else | |
521 | ec->cond_met = 0; | |
522 | ||
49bb9e6d | 523 | /* Non-Linear Processing */ |
4460a860 M |
524 | |
525 | ec->clean_nlp = ec->clean; | |
526 | if (ec->adaption_mode & ECHO_CAN_USE_NLP) { | |
49bb9e6d GKH |
527 | /* |
528 | * Non-linear processor - a fancy way to say "zap small | |
529 | * signals, to avoid residual echo due to (uLaw/ALaw) | |
530 | * non-linearity in the channel.". | |
531 | */ | |
4460a860 M |
532 | |
533 | if ((16 * ec->Lclean < ec->Ltx)) { | |
49bb9e6d GKH |
534 | /* |
535 | * Our e/c has improved echo by at least 24 dB (each | |
536 | * factor of 2 is 6dB, so 2*2*2*2=16 is the same as | |
537 | * 6+6+6+6=24dB) | |
538 | */ | |
4460a860 M |
539 | if (ec->adaption_mode & ECHO_CAN_USE_CNG) { |
540 | ec->cng_level = ec->Lbgn; | |
541 | ||
49bb9e6d GKH |
542 | /* |
543 | * Very elementary comfort noise generation. | |
544 | * Just random numbers rolled off very vaguely | |
545 | * Hoth-like. DR: This noise doesn't sound | |
546 | * quite right to me - I suspect there are some | |
83aa3c7b | 547 | * overflow issues in the filtering as it's too |
49bb9e6d GKH |
548 | * "crackly". |
549 | * TODO: debug this, maybe just play noise at | |
550 | * high level or look at spectrum. | |
4460a860 M |
551 | */ |
552 | ||
553 | ec->cng_rndnum = | |
554 | 1664525U * ec->cng_rndnum + 1013904223U; | |
555 | ec->cng_filter = | |
556 | ((ec->cng_rndnum & 0xFFFF) - 32768 + | |
557 | 5 * ec->cng_filter) >> 3; | |
558 | ec->clean_nlp = | |
559 | (ec->cng_filter * ec->cng_level * 8) >> 14; | |
560 | ||
561 | } else if (ec->adaption_mode & ECHO_CAN_USE_CLIP) { | |
562 | /* This sounds much better than CNG */ | |
563 | if (ec->clean_nlp > ec->Lbgn) | |
564 | ec->clean_nlp = ec->Lbgn; | |
565 | if (ec->clean_nlp < -ec->Lbgn) | |
566 | ec->clean_nlp = -ec->Lbgn; | |
567 | } else { | |
49bb9e6d GKH |
568 | /* |
569 | * just mute the residual, doesn't sound very | |
570 | * good, used mainly in G168 tests | |
571 | */ | |
4460a860 M |
572 | ec->clean_nlp = 0; |
573 | } | |
574 | } else { | |
49bb9e6d GKH |
575 | /* |
576 | * Background noise estimator. I tried a few | |
577 | * algorithms here without much luck. This very simple | |
578 | * one seems to work best, we just average the level | |
579 | * using a slow (1 sec time const) filter if the | |
580 | * current level is less than a (experimentally | |
581 | * derived) constant. This means we dont include high | |
582 | * level signals like near end speech. When combined | |
583 | * with CNG or especially CLIP seems to work OK. | |
4460a860 M |
584 | */ |
585 | if (ec->Lclean < 40) { | |
586 | ec->Lbgn_acc += abs(ec->clean) - ec->Lbgn; | |
587 | ec->Lbgn = (ec->Lbgn_acc + (1 << 11)) >> 12; | |
588 | } | |
589 | } | |
590 | } | |
591 | ||
592 | /* Roll around the taps buffer */ | |
593 | if (ec->curr_pos <= 0) | |
594 | ec->curr_pos = ec->taps; | |
595 | ec->curr_pos--; | |
596 | ||
597 | if (ec->adaption_mode & ECHO_CAN_DISABLE) | |
598 | ec->clean_nlp = rx; | |
599 | ||
600 | /* Output scaled back up again to match input scaling */ | |
601 | ||
602 | return (int16_t) ec->clean_nlp << 1; | |
10602db8 | 603 | } |
9d8f2d5d | 604 | EXPORT_SYMBOL_GPL(oslec_update); |
10602db8 | 605 | |
935e99fb | 606 | /* This function is separated from the echo canceller is it is usually called |
10602db8 DR |
607 | as part of the tx process. See rx HP (DC blocking) filter above, it's |
608 | the same design. | |
609 | ||
610 | Some soft phones send speech signals with a lot of low frequency | |
611 | energy, e.g. down to 20Hz. This can make the hybrid non-linear | |
612 | which causes the echo canceller to fall over. This filter can help | |
613 | by removing any low frequency before it gets to the tx port of the | |
614 | hybrid. | |
615 | ||
616 | It can also help by removing and DC in the tx signal. DC is bad | |
617 | for LMS algorithms. | |
618 | ||
49bb9e6d GKH |
619 | This is one of the classic DC removal filters, adjusted to provide |
620 | sufficient bass rolloff to meet the above requirement to protect hybrids | |
621 | from things that upset them. The difference between successive samples | |
622 | produces a lousy HPF, and then a suitably placed pole flattens things out. | |
623 | The final result is a nicely rolled off bass end. The filtering is | |
624 | implemented with extended fractional precision, which noise shapes things, | |
625 | giving very clean DC removal. | |
10602db8 DR |
626 | */ |
627 | ||
dc57a3ea | 628 | int16_t oslec_hpf_tx(struct oslec_state *ec, int16_t tx) |
4460a860 M |
629 | { |
630 | int tmp, tmp1; | |
10602db8 | 631 | |
4460a860 M |
632 | if (ec->adaption_mode & ECHO_CAN_USE_TX_HPF) { |
633 | tmp = tx << 15; | |
196e76e8 | 634 | |
49bb9e6d GKH |
635 | /* |
636 | * Make sure the gain of the HPF is 1.0. The first can still | |
637 | * saturate a little under impulse conditions, and it might | |
638 | * roll to 32768 and need clipping on sustained peak level | |
639 | * signals. However, the scale of such clipping is small, and | |
640 | * the error due to any saturation should not markedly affect | |
641 | * the downstream processing. | |
642 | */ | |
4460a860 | 643 | tmp -= (tmp >> 4); |
196e76e8 | 644 | |
4460a860 M |
645 | ec->tx_1 += -(ec->tx_1 >> DC_LOG2BETA) + tmp - ec->tx_2; |
646 | tmp1 = ec->tx_1 >> 15; | |
647 | if (tmp1 > 32767) | |
648 | tmp1 = 32767; | |
649 | if (tmp1 < -32767) | |
650 | tmp1 = -32767; | |
651 | tx = tmp1; | |
652 | ec->tx_2 = tmp; | |
653 | } | |
654 | ||
655 | return tx; | |
10602db8 | 656 | } |
9d8f2d5d | 657 | EXPORT_SYMBOL_GPL(oslec_hpf_tx); |
68b8d9f6 TC |
658 | |
659 | MODULE_LICENSE("GPL"); | |
660 | MODULE_AUTHOR("David Rowe"); | |
661 | MODULE_DESCRIPTION("Open Source Line Echo Canceller"); | |
662 | MODULE_VERSION("0.3.0"); |