Tristan Matthews | 0a329cc | 2013-07-17 13:20:14 -0400 | [diff] [blame] | 1 | /* $Id$ */ |
| 2 | /* |
| 3 | * Copyright (C) 2008-2011 Teluu Inc. (http://www.teluu.com) |
| 4 | * Copyright (C) 2003-2008 Benny Prijono <benny@prijono.org> |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License as published by |
| 8 | * the Free Software Foundation; either version 2 of the License, or |
| 9 | * (at your option) any later version. |
| 10 | * |
| 11 | * This program is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | * GNU General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU General Public License |
| 17 | * along with this program; if not, write to the Free Software |
| 18 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 19 | */ |
| 20 | #include <pjmedia/types.h> |
| 21 | #include <pjmedia/alaw_ulaw.h> |
| 22 | #include <pjmedia/errno.h> |
| 23 | #include <pjmedia/frame.h> |
| 24 | #include <pjmedia/silencedet.h> |
| 25 | #include <pj/array.h> |
| 26 | #include <pj/assert.h> |
| 27 | #include <pj/lock.h> |
| 28 | #include <pj/log.h> |
| 29 | #include <pj/os.h> |
| 30 | #include <pj/pool.h> |
| 31 | |
| 32 | #include "echo_internal.h" |
| 33 | |
| 34 | #define THIS_FILE "echo_suppress.c" |
| 35 | |
| 36 | /* Maximum float constant */ |
| 37 | #define MAX_FLOAT (float)1.701411e38 |
| 38 | |
| 39 | /* The effective learn duration (in seconds) before we declare that learning |
| 40 | * is complete. The actual learning duration itself may be longer depending |
| 41 | * on the conversation pattern (e.g. we can't detect echo if speaker is only |
| 42 | * playing silence). |
| 43 | */ |
| 44 | #define MAX_CALC_DURATION_SEC 3 |
| 45 | |
| 46 | /* The internal audio segment length, in milliseconds. 10ms shold be good |
| 47 | * and no need to change it. |
| 48 | */ |
| 49 | #define SEGMENT_PTIME 10 |
| 50 | |
| 51 | /* The length of the template signal in milliseconds. The longer the template, |
| 52 | * the better correlation will be found, at the expense of more processing |
| 53 | * and longer learning time. |
| 54 | */ |
| 55 | #define TEMPLATE_PTIME 200 |
| 56 | |
| 57 | /* How long to look back in the past to see if either mic or speaker is |
| 58 | * active. |
| 59 | */ |
| 60 | #define SIGNAL_LOOKUP_MSEC 200 |
| 61 | |
| 62 | /* The minimum level value to be considered as talking, in uLaw complement |
| 63 | * (0-255). |
| 64 | */ |
| 65 | #define MIN_SIGNAL_ULAW 35 |
| 66 | |
| 67 | /* The period (in seconds) on which the ES will analize it's effectiveness, |
| 68 | * and it may trigger soft-reset to force recalculation. |
| 69 | */ |
| 70 | #define CHECK_PERIOD 30 |
| 71 | |
| 72 | /* Maximum signal level of average echo residue (in uLaw complement). When |
| 73 | * the residue value exceeds this value, we force the ES to re-learn. |
| 74 | */ |
| 75 | #define MAX_RESIDUE 2.5 |
| 76 | |
| 77 | |
| 78 | #if 0 |
| 79 | # define TRACE_(expr) PJ_LOG(5,expr) |
| 80 | #else |
| 81 | # define TRACE_(expr) |
| 82 | #endif |
| 83 | |
| 84 | PJ_INLINE(float) FABS(float val) |
| 85 | { |
| 86 | if (val < 0) |
| 87 | return -val; |
| 88 | else |
| 89 | return val; |
| 90 | } |
| 91 | |
| 92 | |
| 93 | #if defined(PJ_HAS_FLOATING_POINT) && PJ_HAS_FLOATING_POINT!=0 |
| 94 | typedef float pj_ufloat_t; |
| 95 | # define pj_ufloat_from_float(f) (f) |
| 96 | # define pj_ufloat_mul_u(val1, f) ((val1) * (f)) |
| 97 | # define pj_ufloat_mul_i(val1, f) ((val1) * (f)) |
| 98 | #else |
| 99 | typedef pj_uint32_t pj_ufloat_t; |
| 100 | |
| 101 | pj_ufloat_t pj_ufloat_from_float(float f) |
| 102 | { |
| 103 | return (pj_ufloat_t)(f * 65536); |
| 104 | } |
| 105 | |
| 106 | unsigned pj_ufloat_mul_u(unsigned val1, pj_ufloat_t val2) |
| 107 | { |
| 108 | return (val1 * val2) >> 16; |
| 109 | } |
| 110 | |
| 111 | int pj_ufloat_mul_i(int val1, pj_ufloat_t val2) |
| 112 | { |
| 113 | return (val1 * (pj_int32_t)val2) >> 16; |
| 114 | } |
| 115 | #endif |
| 116 | |
| 117 | |
| 118 | /* Conversation state */ |
| 119 | typedef enum talk_state |
| 120 | { |
| 121 | ST_NULL, |
| 122 | ST_LOCAL_TALK, |
| 123 | ST_REM_SILENT, |
| 124 | ST_DOUBLETALK, |
| 125 | ST_REM_TALK |
| 126 | } talk_state_t; |
| 127 | |
| 128 | const char *state_names[] = |
| 129 | { |
| 130 | "Null", |
| 131 | "local talking", |
| 132 | "remote silent", |
| 133 | "doubletalk", |
| 134 | "remote talking" |
| 135 | }; |
| 136 | |
| 137 | |
| 138 | /* Description: |
| 139 | |
| 140 | The echo suppressor tries to find the position of echoed signal by looking |
| 141 | at the correlation between signal played to the speaker (played signal) |
| 142 | and the signal captured from the microphone (recorded signal). |
| 143 | |
| 144 | To do this, it first divides the frames (from mic and speaker) into |
| 145 | segments, calculate the audio level of the segment, and save the level |
| 146 | information in the playback and record history (play_hist and rec_hist |
| 147 | respectively). |
| 148 | |
| 149 | In the history, the newest element (depicted as "t0" in the diagram belo) |
| 150 | is put in the last position of the array. |
| 151 | |
| 152 | The record history size is as large as the template size (tmpl_cnt), since |
| 153 | we will use the record history as the template to find the best matching |
| 154 | position in the playback history. |
| 155 | |
| 156 | Here is the record history buffer: |
| 157 | |
| 158 | <--templ_cnt--> |
| 159 | +-------------+ |
| 160 | | rec_hist | |
| 161 | +-------------+ |
| 162 | t-templ_cnt......t0 |
| 163 | |
| 164 | As you can see, the newest frame ("t0") is put as the last element in the |
| 165 | array. |
| 166 | |
| 167 | The playback history size is larger than record history, since we need to |
| 168 | find the matching pattern in the past. The playback history size is |
| 169 | "templ_cnt + tail_cnt", where "tail_cnt" is the number of segments equal |
| 170 | to the maximum tail length. The maximum tail length is set when the ES |
| 171 | is created. |
| 172 | |
| 173 | Here is the playback history buffer: |
| 174 | |
| 175 | <-----tail_cnt-----> <--templ_cnt--> |
| 176 | +-------------------+--------------+ |
| 177 | | play_hist | |
| 178 | +-------------------+--------------+ |
| 179 | t-play_hist_cnt...t-templ_cnt.......t0 |
| 180 | |
| 181 | |
| 182 | |
| 183 | Learning: |
| 184 | |
| 185 | During the processing, the ES calculates the following values: |
| 186 | - the correlation value, that is how similar the playback signal compared |
| 187 | to the mic signal. The lower the correlation value the better (i.e. more |
| 188 | similar) the signal is. The correlation value is done over the template |
| 189 | duration. |
| 190 | - the gain scaling factor, that is the ratio between mic signal and |
| 191 | speaker signal. The ES calculates both the minimum and average ratios. |
| 192 | |
| 193 | The ES calculates both the values above for every tail position in the |
| 194 | playback history. The values are saved in arrays below: |
| 195 | |
| 196 | <-----tail_cnt-----> |
| 197 | +-------------------+ |
| 198 | | corr_sum | |
| 199 | +-------------------+ |
| 200 | | min_factor | |
| 201 | +-------------------+ |
| 202 | | avg_factor | |
| 203 | +-------------------+ |
| 204 | |
| 205 | At the end of processing, the ES iterates through the correlation array and |
| 206 | picks the tail index with the lowest corr_sum value. This is the position |
| 207 | where echo is most likely to be found. |
| 208 | |
| 209 | |
| 210 | Processing: |
| 211 | |
| 212 | Once learning is done, the ES will change the level of the mic signal |
| 213 | depending on the state of the conversation and according to the ratio that |
| 214 | has been found in the learning phase above. |
| 215 | |
| 216 | */ |
| 217 | |
| 218 | /* |
| 219 | * The simple echo suppresor state |
| 220 | */ |
| 221 | typedef struct echo_supp |
| 222 | { |
| 223 | unsigned clock_rate; /* Clock rate. */ |
| 224 | pj_uint16_t samples_per_frame; /* Frame length in samples */ |
| 225 | pj_uint16_t samples_per_segment;/* Segment length in samples */ |
| 226 | pj_uint16_t tail_ms; /* Tail length in milliseconds */ |
| 227 | pj_uint16_t tail_samples; /* Tail length in samples. */ |
| 228 | |
| 229 | pj_bool_t learning; /* Are we still learning yet? */ |
| 230 | talk_state_t talk_state; /* Current talking state */ |
| 231 | int tail_index; /* Echo location, -1 if not found */ |
| 232 | |
| 233 | unsigned max_calc; /* # of calc before learning complete. |
| 234 | (see MAX_CALC_DURATION_SEC) */ |
| 235 | unsigned calc_cnt; /* Number of calculations so far */ |
| 236 | |
| 237 | unsigned update_cnt; /* # of updates */ |
| 238 | unsigned templ_cnt; /* Template length, in # of segments */ |
| 239 | unsigned tail_cnt; /* Tail length, in # of segments */ |
| 240 | unsigned play_hist_cnt; /* # of segments in play_hist */ |
| 241 | pj_uint16_t *play_hist; /* Array of playback levels */ |
| 242 | pj_uint16_t *rec_hist; /* Array of rec levels */ |
| 243 | |
| 244 | float *corr_sum; /* Array of corr for each tail pos. */ |
| 245 | float *tmp_corr; /* Temporary corr array calculation */ |
| 246 | float best_corr; /* Best correlation so far. */ |
| 247 | |
| 248 | unsigned sum_rec_level; /* Running sum of level in rec_hist */ |
| 249 | float rec_corr; /* Running corr in rec_hist. */ |
| 250 | |
| 251 | unsigned sum_play_level0; /* Running sum of level for first pos */ |
| 252 | float play_corr0; /* Running corr for first pos . */ |
| 253 | |
| 254 | float *min_factor; /* Array of minimum scaling factor */ |
| 255 | float *avg_factor; /* Array of average scaling factor */ |
| 256 | float *tmp_factor; /* Array to store provisional result */ |
| 257 | |
| 258 | unsigned running_cnt; /* Running duration in # of frames */ |
| 259 | float residue; /* Accummulated echo residue. */ |
| 260 | float last_factor; /* Last factor applied to mic signal */ |
| 261 | } echo_supp; |
| 262 | |
| 263 | |
| 264 | |
| 265 | /* |
| 266 | * Create. |
| 267 | */ |
| 268 | PJ_DEF(pj_status_t) echo_supp_create( pj_pool_t *pool, |
| 269 | unsigned clock_rate, |
| 270 | unsigned channel_count, |
| 271 | unsigned samples_per_frame, |
| 272 | unsigned tail_ms, |
| 273 | unsigned options, |
| 274 | void **p_state ) |
| 275 | { |
| 276 | echo_supp *ec; |
| 277 | |
| 278 | PJ_UNUSED_ARG(channel_count); |
| 279 | PJ_UNUSED_ARG(options); |
| 280 | |
| 281 | PJ_ASSERT_RETURN(samples_per_frame >= SEGMENT_PTIME * clock_rate / 1000, |
| 282 | PJ_ENOTSUP); |
| 283 | |
| 284 | ec = PJ_POOL_ZALLOC_T(pool, struct echo_supp); |
| 285 | ec->clock_rate = clock_rate; |
| 286 | ec->samples_per_frame = (pj_uint16_t)samples_per_frame; |
| 287 | ec->samples_per_segment = (pj_uint16_t)(SEGMENT_PTIME * clock_rate / 1000); |
| 288 | ec->tail_ms = (pj_uint16_t)tail_ms; |
| 289 | ec->tail_samples = (pj_uint16_t)(tail_ms * clock_rate / 1000); |
| 290 | |
| 291 | ec->templ_cnt = TEMPLATE_PTIME / SEGMENT_PTIME; |
| 292 | ec->tail_cnt = (pj_uint16_t)(tail_ms / SEGMENT_PTIME); |
| 293 | ec->play_hist_cnt = (pj_uint16_t)(ec->tail_cnt+ec->templ_cnt); |
| 294 | |
| 295 | ec->max_calc = (pj_uint16_t)(MAX_CALC_DURATION_SEC * clock_rate / |
| 296 | ec->samples_per_segment); |
| 297 | |
| 298 | ec->rec_hist = (pj_uint16_t*) |
| 299 | pj_pool_alloc(pool, ec->templ_cnt * |
| 300 | sizeof(ec->rec_hist[0])); |
| 301 | |
| 302 | /* Note: play history has twice number of elements */ |
| 303 | ec->play_hist = (pj_uint16_t*) |
| 304 | pj_pool_alloc(pool, ec->play_hist_cnt * |
| 305 | sizeof(ec->play_hist[0])); |
| 306 | |
| 307 | ec->corr_sum = (float*) |
| 308 | pj_pool_alloc(pool, ec->tail_cnt * |
| 309 | sizeof(ec->corr_sum[0])); |
| 310 | ec->tmp_corr = (float*) |
| 311 | pj_pool_alloc(pool, ec->tail_cnt * |
| 312 | sizeof(ec->tmp_corr[0])); |
| 313 | ec->min_factor = (float*) |
| 314 | pj_pool_alloc(pool, ec->tail_cnt * |
| 315 | sizeof(ec->min_factor[0])); |
| 316 | ec->avg_factor = (float*) |
| 317 | pj_pool_alloc(pool, ec->tail_cnt * |
| 318 | sizeof(ec->avg_factor[0])); |
| 319 | ec->tmp_factor = (float*) |
| 320 | pj_pool_alloc(pool, ec->tail_cnt * |
| 321 | sizeof(ec->tmp_factor[0])); |
| 322 | echo_supp_reset(ec); |
| 323 | |
| 324 | *p_state = ec; |
| 325 | return PJ_SUCCESS; |
| 326 | } |
| 327 | |
| 328 | |
| 329 | /* |
| 330 | * Destroy. |
| 331 | */ |
| 332 | PJ_DEF(pj_status_t) echo_supp_destroy(void *state) |
| 333 | { |
| 334 | PJ_UNUSED_ARG(state); |
| 335 | return PJ_SUCCESS; |
| 336 | } |
| 337 | |
| 338 | |
| 339 | /* |
| 340 | * Hard reset |
| 341 | */ |
| 342 | PJ_DEF(void) echo_supp_reset(void *state) |
| 343 | { |
| 344 | unsigned i; |
| 345 | echo_supp *ec = (echo_supp*) state; |
| 346 | |
| 347 | pj_bzero(ec->rec_hist, ec->templ_cnt * sizeof(ec->rec_hist[0])); |
| 348 | pj_bzero(ec->play_hist, ec->play_hist_cnt * sizeof(ec->play_hist[0])); |
| 349 | |
| 350 | for (i=0; i<ec->tail_cnt; ++i) { |
| 351 | ec->corr_sum[i] = ec->avg_factor[i] = 0; |
| 352 | ec->min_factor[i] = MAX_FLOAT; |
| 353 | } |
| 354 | |
| 355 | ec->update_cnt = 0; |
| 356 | ec->calc_cnt = 0; |
| 357 | ec->learning = PJ_TRUE; |
| 358 | ec->tail_index = -1; |
| 359 | ec->best_corr = MAX_FLOAT; |
| 360 | ec->talk_state = ST_NULL; |
| 361 | ec->last_factor = 1.0; |
| 362 | ec->residue = 0; |
| 363 | ec->running_cnt = 0; |
| 364 | ec->sum_rec_level = ec->sum_play_level0 = 0; |
| 365 | ec->rec_corr = ec->play_corr0 = 0; |
| 366 | } |
| 367 | |
| 368 | /* |
| 369 | * Soft reset to force the EC to re-learn without having to discard all |
| 370 | * rec and playback history. |
| 371 | */ |
| 372 | PJ_DEF(void) echo_supp_soft_reset(void *state) |
| 373 | { |
| 374 | unsigned i; |
| 375 | |
| 376 | echo_supp *ec = (echo_supp*) state; |
| 377 | |
| 378 | for (i=0; i<ec->tail_cnt; ++i) { |
| 379 | ec->corr_sum[i] = 0; |
| 380 | } |
| 381 | |
| 382 | ec->update_cnt = 0; |
| 383 | ec->calc_cnt = 0; |
| 384 | ec->learning = PJ_TRUE; |
| 385 | ec->best_corr = MAX_FLOAT; |
| 386 | ec->residue = 0; |
| 387 | ec->running_cnt = 0; |
| 388 | ec->sum_rec_level = ec->sum_play_level0 = 0; |
| 389 | ec->rec_corr = ec->play_corr0 = 0; |
| 390 | |
| 391 | PJ_LOG(4,(THIS_FILE, "Echo suppressor soft reset. Re-learning..")); |
| 392 | } |
| 393 | |
| 394 | |
| 395 | /* Set state */ |
| 396 | static void echo_supp_set_state(echo_supp *ec, talk_state_t state, |
| 397 | unsigned level) |
| 398 | { |
| 399 | PJ_UNUSED_ARG(level); |
| 400 | |
| 401 | if (state != ec->talk_state) { |
| 402 | TRACE_((THIS_FILE, "[%03d.%03d] %s --> %s, level=%u", |
| 403 | (ec->update_cnt * SEGMENT_PTIME / 1000), |
| 404 | ((ec->update_cnt * SEGMENT_PTIME) % 1000), |
| 405 | state_names[ec->talk_state], |
| 406 | state_names[state], level)); |
| 407 | ec->talk_state = state; |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | /* |
| 412 | * Update EC state |
| 413 | */ |
| 414 | static void echo_supp_update(echo_supp *ec, pj_int16_t *rec_frm, |
| 415 | const pj_int16_t *play_frm) |
| 416 | { |
| 417 | int prev_index; |
| 418 | unsigned i, j, frm_level, sum_play_level, ulaw; |
| 419 | pj_uint16_t old_rec_frm_level, old_play_frm_level; |
| 420 | float play_corr; |
| 421 | |
| 422 | ++ec->update_cnt; |
| 423 | if (ec->update_cnt > 0x7FFFFFFF) |
| 424 | ec->update_cnt = 0x7FFFFFFF; /* Detect overflow */ |
| 425 | |
| 426 | /* Calculate current play frame level */ |
| 427 | frm_level = pjmedia_calc_avg_signal(play_frm, ec->samples_per_segment); |
| 428 | ++frm_level; /* to avoid division by zero */ |
| 429 | |
| 430 | /* Save the oldest frame level for later */ |
| 431 | old_play_frm_level = ec->play_hist[0]; |
| 432 | |
| 433 | /* Push current frame level to the back of the play history */ |
| 434 | pj_array_erase(ec->play_hist, sizeof(pj_uint16_t), ec->play_hist_cnt, 0); |
| 435 | ec->play_hist[ec->play_hist_cnt-1] = (pj_uint16_t) frm_level; |
| 436 | |
| 437 | /* Calculate level of current mic frame */ |
| 438 | frm_level = pjmedia_calc_avg_signal(rec_frm, ec->samples_per_segment); |
| 439 | ++frm_level; /* to avoid division by zero */ |
| 440 | |
| 441 | /* Save the oldest frame level for later */ |
| 442 | old_rec_frm_level = ec->rec_hist[0]; |
| 443 | |
| 444 | /* Push to the back of the rec history */ |
| 445 | pj_array_erase(ec->rec_hist, sizeof(pj_uint16_t), ec->templ_cnt, 0); |
| 446 | ec->rec_hist[ec->templ_cnt-1] = (pj_uint16_t) frm_level; |
| 447 | |
| 448 | |
| 449 | /* Can't do the calc until the play history is full. */ |
| 450 | if (ec->update_cnt < ec->play_hist_cnt) |
| 451 | return; |
| 452 | |
| 453 | /* Skip if learning is done */ |
| 454 | if (!ec->learning) |
| 455 | return; |
| 456 | |
| 457 | |
| 458 | /* Calculate rec signal pattern */ |
| 459 | if (ec->sum_rec_level == 0) { |
| 460 | /* Buffer has just been filled up, do full calculation */ |
| 461 | ec->rec_corr = 0; |
| 462 | ec->sum_rec_level = 0; |
| 463 | for (i=0; i < ec->templ_cnt-1; ++i) { |
| 464 | float corr; |
| 465 | corr = (float)ec->rec_hist[i+1] / ec->rec_hist[i]; |
| 466 | ec->rec_corr += corr; |
| 467 | ec->sum_rec_level += ec->rec_hist[i]; |
| 468 | } |
| 469 | ec->sum_rec_level += ec->rec_hist[i]; |
| 470 | } else { |
| 471 | /* Update from previous calculation */ |
| 472 | ec->sum_rec_level = ec->sum_rec_level - old_rec_frm_level + |
| 473 | ec->rec_hist[ec->templ_cnt-1]; |
| 474 | ec->rec_corr = ec->rec_corr - ((float)ec->rec_hist[0] / |
| 475 | old_rec_frm_level) + |
| 476 | ((float)ec->rec_hist[ec->templ_cnt-1] / |
| 477 | ec->rec_hist[ec->templ_cnt-2]); |
| 478 | } |
| 479 | |
| 480 | /* Iterate through the play history and calculate the signal correlation |
| 481 | * for every tail position in the play_hist. Save the result in temporary |
| 482 | * array since we may bail out early if the conversation state is not good |
| 483 | * to detect echo. |
| 484 | */ |
| 485 | /* |
| 486 | * First phase: do full calculation for the first position |
| 487 | */ |
| 488 | if (ec->sum_play_level0 == 0) { |
| 489 | /* Buffer has just been filled up, do full calculation */ |
| 490 | sum_play_level = 0; |
| 491 | play_corr = 0; |
| 492 | for (j=0; j<ec->templ_cnt-1; ++j) { |
| 493 | float corr; |
| 494 | corr = (float)ec->play_hist[j+1] / ec->play_hist[j]; |
| 495 | play_corr += corr; |
| 496 | sum_play_level += ec->play_hist[j]; |
| 497 | } |
| 498 | sum_play_level += ec->play_hist[j]; |
| 499 | ec->sum_play_level0 = sum_play_level; |
| 500 | ec->play_corr0 = play_corr; |
| 501 | } else { |
| 502 | /* Update from previous calculation */ |
| 503 | ec->sum_play_level0 = ec->sum_play_level0 - old_play_frm_level + |
| 504 | ec->play_hist[ec->templ_cnt-1]; |
| 505 | ec->play_corr0 = ec->play_corr0 - ((float)ec->play_hist[0] / |
| 506 | old_play_frm_level) + |
| 507 | ((float)ec->play_hist[ec->templ_cnt-1] / |
| 508 | ec->play_hist[ec->templ_cnt-2]); |
| 509 | sum_play_level = ec->sum_play_level0; |
| 510 | play_corr = ec->play_corr0; |
| 511 | } |
| 512 | ec->tmp_corr[0] = FABS(play_corr - ec->rec_corr); |
| 513 | ec->tmp_factor[0] = (float)ec->sum_rec_level / sum_play_level; |
| 514 | |
| 515 | /* Bail out if remote isn't talking */ |
| 516 | ulaw = pjmedia_linear2ulaw(sum_play_level/ec->templ_cnt) ^ 0xFF; |
| 517 | if (ulaw < MIN_SIGNAL_ULAW) { |
| 518 | echo_supp_set_state(ec, ST_REM_SILENT, ulaw); |
| 519 | return; |
| 520 | } |
| 521 | /* Bail out if local user is talking */ |
| 522 | if (ec->sum_rec_level >= sum_play_level) { |
| 523 | echo_supp_set_state(ec, ST_LOCAL_TALK, ulaw); |
| 524 | return; |
| 525 | } |
| 526 | |
| 527 | /* |
| 528 | * Second phase: do incremental calculation for the rest of positions |
| 529 | */ |
| 530 | for (i=1; i < ec->tail_cnt; ++i) { |
| 531 | unsigned end; |
| 532 | |
| 533 | end = i + ec->templ_cnt; |
| 534 | |
| 535 | sum_play_level = sum_play_level - ec->play_hist[i-1] + |
| 536 | ec->play_hist[end-1]; |
| 537 | play_corr = play_corr - ((float)ec->play_hist[i]/ec->play_hist[i-1]) + |
| 538 | ((float)ec->play_hist[end-1]/ec->play_hist[end-2]); |
| 539 | |
| 540 | /* Bail out if remote isn't talking */ |
| 541 | ulaw = pjmedia_linear2ulaw(sum_play_level/ec->templ_cnt) ^ 0xFF; |
| 542 | if (ulaw < MIN_SIGNAL_ULAW) { |
| 543 | echo_supp_set_state(ec, ST_REM_SILENT, ulaw); |
| 544 | return; |
| 545 | } |
| 546 | |
| 547 | /* Bail out if local user is talking */ |
| 548 | if (ec->sum_rec_level >= sum_play_level) { |
| 549 | echo_supp_set_state(ec, ST_LOCAL_TALK, ulaw); |
| 550 | return; |
| 551 | } |
| 552 | |
| 553 | #if 0 |
| 554 | // disabled: not a good idea if mic throws out loud echo |
| 555 | /* Also bail out if we suspect there's a doubletalk */ |
| 556 | ulaw = pjmedia_linear2ulaw(ec->sum_rec_level/ec->templ_cnt) ^ 0xFF; |
| 557 | if (ulaw > MIN_SIGNAL_ULAW) { |
| 558 | echo_supp_set_state(ec, ST_DOUBLETALK, ulaw); |
| 559 | return; |
| 560 | } |
| 561 | #endif |
| 562 | |
| 563 | /* Calculate correlation and save to temporary array */ |
| 564 | ec->tmp_corr[i] = FABS(play_corr - ec->rec_corr); |
| 565 | |
| 566 | /* Also calculate the gain factor between mic and speaker level */ |
| 567 | ec->tmp_factor[i] = (float)ec->sum_rec_level / sum_play_level; |
| 568 | pj_assert(ec->tmp_factor[i] < 1); |
| 569 | } |
| 570 | |
| 571 | /* We seem to have good signal, we can update the EC state */ |
| 572 | echo_supp_set_state(ec, ST_REM_TALK, MIN_SIGNAL_ULAW); |
| 573 | |
| 574 | /* Accummulate the correlation value to the history and at the same |
| 575 | * time find the tail index of the best correlation. |
| 576 | */ |
| 577 | prev_index = ec->tail_index; |
| 578 | for (i=1; i<ec->tail_cnt-1; ++i) { |
| 579 | float *p = &ec->corr_sum[i], sum; |
| 580 | |
| 581 | /* Accummulate correlation value for this tail position */ |
| 582 | ec->corr_sum[i] += ec->tmp_corr[i]; |
| 583 | |
| 584 | /* Update the min and avg gain factor for this tail position */ |
| 585 | if (ec->tmp_factor[i] < ec->min_factor[i]) |
| 586 | ec->min_factor[i] = ec->tmp_factor[i]; |
| 587 | ec->avg_factor[i] = ((ec->avg_factor[i] * ec->tail_cnt) + |
| 588 | ec->tmp_factor[i]) / |
| 589 | (ec->tail_cnt + 1); |
| 590 | |
| 591 | /* To get the best correlation, also include the correlation |
| 592 | * value of the neighbouring tail locations. |
| 593 | */ |
| 594 | sum = *(p-1) + (*p)*2 + *(p+1); |
| 595 | //sum = *p; |
| 596 | |
| 597 | /* See if we have better correlation value */ |
| 598 | if (sum < ec->best_corr) { |
| 599 | ec->tail_index = i; |
| 600 | ec->best_corr = sum; |
| 601 | } |
| 602 | } |
| 603 | |
| 604 | if (ec->tail_index != prev_index) { |
| 605 | unsigned duration; |
| 606 | int imin, iavg; |
| 607 | |
| 608 | duration = ec->update_cnt * SEGMENT_PTIME; |
| 609 | imin = (int)(ec->min_factor[ec->tail_index] * 1000); |
| 610 | iavg = (int)(ec->avg_factor[ec->tail_index] * 1000); |
| 611 | |
| 612 | PJ_LOG(4,(THIS_FILE, |
| 613 | "Echo suppressor updated at t=%03d.%03ds, echo tail=%d msec" |
| 614 | ", factor min/avg=%d.%03d/%d.%03d", |
| 615 | (duration/1000), (duration%1000), |
| 616 | (ec->tail_cnt-ec->tail_index) * SEGMENT_PTIME, |
| 617 | imin/1000, imin%1000, |
| 618 | iavg/1000, iavg%1000)); |
| 619 | |
| 620 | } |
| 621 | |
| 622 | ++ec->calc_cnt; |
| 623 | |
| 624 | if (ec->calc_cnt > ec->max_calc) { |
| 625 | unsigned duration; |
| 626 | int imin, iavg; |
| 627 | |
| 628 | |
| 629 | ec->learning = PJ_FALSE; |
| 630 | ec->running_cnt = 0; |
| 631 | |
| 632 | duration = ec->update_cnt * SEGMENT_PTIME; |
| 633 | imin = (int)(ec->min_factor[ec->tail_index] * 1000); |
| 634 | iavg = (int)(ec->avg_factor[ec->tail_index] * 1000); |
| 635 | |
| 636 | PJ_LOG(4,(THIS_FILE, |
| 637 | "Echo suppressor learning done at t=%03d.%03ds, tail=%d ms" |
| 638 | ", factor min/avg=%d.%03d/%d.%03d", |
| 639 | (duration/1000), (duration%1000), |
| 640 | (ec->tail_cnt-ec->tail_index) * SEGMENT_PTIME, |
| 641 | imin/1000, imin%1000, |
| 642 | iavg/1000, iavg%1000)); |
| 643 | } |
| 644 | |
| 645 | } |
| 646 | |
| 647 | |
| 648 | /* Amplify frame */ |
| 649 | static void amplify_frame(pj_int16_t *frm, unsigned length, |
| 650 | pj_ufloat_t factor) |
| 651 | { |
| 652 | unsigned i; |
| 653 | |
| 654 | for (i=0; i<length; ++i) { |
| 655 | frm[i] = (pj_int16_t)pj_ufloat_mul_i(frm[i], factor); |
| 656 | } |
| 657 | } |
| 658 | |
| 659 | /* |
| 660 | * Perform echo cancellation. |
| 661 | */ |
| 662 | PJ_DEF(pj_status_t) echo_supp_cancel_echo( void *state, |
| 663 | pj_int16_t *rec_frm, |
| 664 | const pj_int16_t *play_frm, |
| 665 | unsigned options, |
| 666 | void *reserved ) |
| 667 | { |
| 668 | unsigned i, N; |
| 669 | echo_supp *ec = (echo_supp*) state; |
| 670 | |
| 671 | PJ_UNUSED_ARG(options); |
| 672 | PJ_UNUSED_ARG(reserved); |
| 673 | |
| 674 | /* Calculate number of segments. This should be okay even if |
| 675 | * samples_per_frame is not a multiply of samples_per_segment, since |
| 676 | * we only calculate level. |
| 677 | */ |
| 678 | N = ec->samples_per_frame / ec->samples_per_segment; |
| 679 | pj_assert(N>0); |
| 680 | for (i=0; i<N; ++i) { |
| 681 | unsigned pos = i * ec->samples_per_segment; |
| 682 | echo_supp_update(ec, rec_frm+pos, play_frm+pos); |
| 683 | } |
| 684 | |
| 685 | if (ec->tail_index < 0) { |
| 686 | /* Not ready */ |
| 687 | } else { |
| 688 | unsigned lookup_cnt, rec_level=0, play_level=0; |
| 689 | unsigned tail_cnt; |
| 690 | float factor; |
| 691 | |
| 692 | /* How many previous segments to lookup */ |
| 693 | lookup_cnt = SIGNAL_LOOKUP_MSEC / SEGMENT_PTIME; |
| 694 | if (lookup_cnt > ec->templ_cnt) |
| 695 | lookup_cnt = ec->templ_cnt; |
| 696 | |
| 697 | /* Lookup in recording history to get maximum mic level, to see |
| 698 | * if local user is currently talking |
| 699 | */ |
| 700 | for (i=ec->templ_cnt - lookup_cnt; i < ec->templ_cnt; ++i) { |
| 701 | if (ec->rec_hist[i] > rec_level) |
| 702 | rec_level = ec->rec_hist[i]; |
| 703 | } |
| 704 | rec_level = pjmedia_linear2ulaw(rec_level) ^ 0xFF; |
| 705 | |
| 706 | /* Calculate the detected tail length, in # of segments */ |
| 707 | tail_cnt = (ec->tail_cnt - ec->tail_index); |
| 708 | |
| 709 | /* Lookup in playback history to get max speaker level, to see |
| 710 | * if remote user is currently talking |
| 711 | */ |
| 712 | for (i=ec->play_hist_cnt -lookup_cnt -tail_cnt; |
| 713 | i<ec->play_hist_cnt-tail_cnt; ++i) |
| 714 | { |
| 715 | if (ec->play_hist[i] > play_level) |
| 716 | play_level = ec->play_hist[i]; |
| 717 | } |
| 718 | play_level = pjmedia_linear2ulaw(play_level) ^ 0xFF; |
| 719 | |
| 720 | if (rec_level >= MIN_SIGNAL_ULAW) { |
| 721 | if (play_level < MIN_SIGNAL_ULAW) { |
| 722 | /* Mic is talking, speaker is idle. Let mic signal pass as is. |
| 723 | */ |
| 724 | factor = 1.0; |
| 725 | echo_supp_set_state(ec, ST_LOCAL_TALK, rec_level); |
| 726 | } else if (rec_level > play_level) { |
| 727 | /* Seems that both are talking. Scale the mic signal |
| 728 | * down a little bit to reduce echo, while allowing both |
| 729 | * parties to talk at the same time. |
| 730 | */ |
| 731 | factor = (float)(ec->avg_factor[ec->tail_index] * 2); |
| 732 | echo_supp_set_state(ec, ST_DOUBLETALK, rec_level); |
| 733 | } else { |
| 734 | /* Speaker is active, but we've picked up large signal in |
| 735 | * the microphone. Assume that this is an echo, so bring |
| 736 | * the level down to minimum too. |
| 737 | */ |
| 738 | factor = ec->min_factor[ec->tail_index] / 2; |
| 739 | echo_supp_set_state(ec, ST_REM_TALK, play_level); |
| 740 | } |
| 741 | } else { |
| 742 | if (play_level < MIN_SIGNAL_ULAW) { |
| 743 | /* Both mic and speaker seems to be idle. Also scale the |
| 744 | * mic signal down with average factor to reduce low power |
| 745 | * echo. |
| 746 | */ |
| 747 | factor = ec->avg_factor[ec->tail_index] * 3 / 2; |
| 748 | echo_supp_set_state(ec, ST_REM_SILENT, rec_level); |
| 749 | } else { |
| 750 | /* Mic is idle, but there's something playing in speaker. |
| 751 | * Scale the mic down to minimum |
| 752 | */ |
| 753 | factor = ec->min_factor[ec->tail_index] / 2; |
| 754 | echo_supp_set_state(ec, ST_REM_TALK, play_level); |
| 755 | } |
| 756 | } |
| 757 | |
| 758 | /* Smoothen the transition */ |
| 759 | if (factor >= ec->last_factor) |
| 760 | factor = (factor + ec->last_factor) / 2; |
| 761 | else |
| 762 | factor = (factor + ec->last_factor*19) / 20; |
| 763 | |
| 764 | /* Amplify frame */ |
| 765 | amplify_frame(rec_frm, ec->samples_per_frame, |
| 766 | pj_ufloat_from_float(factor)); |
| 767 | ec->last_factor = factor; |
| 768 | |
| 769 | if (ec->talk_state == ST_REM_TALK) { |
| 770 | unsigned level, recalc_cnt; |
| 771 | |
| 772 | /* Get the adjusted frame signal level */ |
| 773 | level = pjmedia_calc_avg_signal(rec_frm, ec->samples_per_frame); |
| 774 | level = pjmedia_linear2ulaw(level) ^ 0xFF; |
| 775 | |
| 776 | /* Accumulate average echo residue to see the ES effectiveness */ |
| 777 | ec->residue = ((ec->residue * ec->running_cnt) + level) / |
| 778 | (ec->running_cnt + 1); |
| 779 | |
| 780 | ++ec->running_cnt; |
| 781 | |
| 782 | /* Check if we need to re-learn */ |
| 783 | recalc_cnt = CHECK_PERIOD * ec->clock_rate / ec->samples_per_frame; |
| 784 | if (ec->running_cnt > recalc_cnt) { |
| 785 | int iresidue; |
| 786 | |
| 787 | iresidue = (int)(ec->residue*1000); |
| 788 | |
| 789 | PJ_LOG(5,(THIS_FILE, "Echo suppressor residue = %d.%03d", |
| 790 | iresidue/1000, iresidue%1000)); |
| 791 | |
| 792 | if (ec->residue > MAX_RESIDUE && !ec->learning) { |
| 793 | echo_supp_soft_reset(ec); |
| 794 | ec->residue = 0; |
| 795 | } else { |
| 796 | ec->running_cnt = 0; |
| 797 | ec->residue = 0; |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | } |
| 802 | |
| 803 | return PJ_SUCCESS; |
| 804 | } |
| 805 | |