1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
|
/* aec.cpp
*
* Copyright (C) DFS Deutsche Flugsicherung (2004, 2005).
* All Rights Reserved.
*
* Acoustic Echo Cancellation NLMS-pw algorithm
*
* Version 0.3 filter created with www.dsptutor.freeuk.com
* Version 0.3.1 Allow change of stability parameter delta
* Version 0.4 Leaky Normalized LMS - pre whitening algorithm
*/
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#ifdef HAVE_CONFIG_H
#include <config.h>
#endif
#include <math.h>
#include <string.h>
#include <stdint.h>
#include <pulse/xmalloc.h>
#include "adrian-aec.h"
#ifndef DISABLE_ORC
#include "adrian-aec-orc-gen.h"
#endif
#ifdef __SSE__
#include <xmmintrin.h>
#endif
/* Vector Dot Product */
static REAL dotp(REAL a[], REAL b[])
{
REAL sum0 = 0.0f, sum1 = 0.0f;
int j;
for (j = 0; j < NLMS_LEN; j += 2) {
// optimize: partial loop unrolling
sum0 += a[j] * b[j];
sum1 += a[j + 1] * b[j + 1];
}
return sum0 + sum1;
}
static REAL dotp_sse(REAL a[], REAL b[])
{
#ifdef __SSE__
/* This is taken from speex's inner product implementation */
int j;
REAL sum;
__m128 acc = _mm_setzero_ps();
for (j=0;j<NLMS_LEN;j+=8)
{
acc = _mm_add_ps(acc, _mm_mul_ps(_mm_load_ps(a+j), _mm_loadu_ps(b+j)));
acc = _mm_add_ps(acc, _mm_mul_ps(_mm_load_ps(a+j+4), _mm_loadu_ps(b+j+4)));
}
acc = _mm_add_ps(acc, _mm_movehl_ps(acc, acc));
acc = _mm_add_ss(acc, _mm_shuffle_ps(acc, acc, 0x55));
_mm_store_ss(&sum, acc);
return sum;
#else
return dotp(a, b);
#endif
}
AEC* AEC_init(int RATE, int have_vector)
{
AEC *a = pa_xnew0(AEC, 1);
a->j = NLMS_EXT;
AEC_setambient(a, NoiseFloor);
a->dfast = a->dslow = M75dB_PCM;
a->xfast = a->xslow = M80dB_PCM;
a->gain = 1.0f;
a->Fx = IIR1_init(2000.0f/RATE);
a->Fe = IIR1_init(2000.0f/RATE);
a->cutoff = FIR_HP_300Hz_init();
a->acMic = IIR_HP_init();
a->acSpk = IIR_HP_init();
a->aes_y2 = M0dB;
a->fdwdisplay = -1;
if (have_vector) {
/* Get a 16-byte aligned location */
a->w = (REAL *) (((uintptr_t) a->w_arr) - (((uintptr_t) a->w_arr) % 16) + 16);
a->dotp = dotp_sse;
} else {
/* We don't care about alignment, just use the array as-is */
a->w = a->w_arr;
a->dotp = dotp;
}
return a;
}
void AEC_done(AEC *a) {
pa_assert(a);
pa_xfree(a->Fx);
pa_xfree(a->Fe);
pa_xfree(a->acMic);
pa_xfree(a->acSpk);
pa_xfree(a->cutoff);
pa_xfree(a);
}
// Adrian soft decision DTD
// (Dual Average Near-End to Far-End signal Ratio DTD)
// This algorithm uses exponential smoothing with different
// ageing parameters to get fast and slow near-end and far-end
// signal averages. The ratio of NFRs term
// (dfast / xfast) / (dslow / xslow) is used to compute the stepsize
// A ratio value of 2.5 is mapped to stepsize 0, a ratio of 0 is
// mapped to 1.0 with a limited linear function.
static float AEC_dtd(AEC *a, REAL d, REAL x)
{
float ratio, stepsize;
// fast near-end and far-end average
a->dfast += ALPHAFAST * (fabsf(d) - a->dfast);
a->xfast += ALPHAFAST * (fabsf(x) - a->xfast);
// slow near-end and far-end average
a->dslow += ALPHASLOW * (fabsf(d) - a->dslow);
a->xslow += ALPHASLOW * (fabsf(x) - a->xslow);
if (a->xfast < M70dB_PCM) {
return 0.0f; // no Spk signal
}
if (a->dfast < M70dB_PCM) {
return 0.0f; // no Mic signal
}
// ratio of NFRs
ratio = (a->dfast * a->xslow) / (a->dslow * a->xfast);
// Linear interpolation with clamping at the limits
if (ratio < STEPX1)
stepsize = STEPY1;
else if (ratio > STEPX2)
stepsize = STEPY2;
else
stepsize = STEPY1 + (STEPY2 - STEPY1) * (ratio - STEPX1) / (STEPX2 - STEPX1);
return stepsize;
}
static void AEC_leaky(AEC *a)
// The xfast signal is used to charge the hangover timer to Thold.
// When hangover expires (no Spk signal for some time) the vector w
// is erased. This is my implementation of Leaky NLMS.
{
if (a->xfast >= M70dB_PCM) {
// vector w is valid for hangover Thold time
a->hangover = Thold;
} else {
if (a->hangover > 1) {
--(a->hangover);
} else if (1 == a->hangover) {
--(a->hangover);
// My Leaky NLMS is to erase vector w when hangover expires
memset(a->w_arr, 0, sizeof(a->w_arr));
}
}
}
#if 0
void AEC::openwdisplay() {
// open TCP connection to program wdisplay.tcl
fdwdisplay = socket_async("127.0.0.1", 50999);
};
#endif
static REAL AEC_nlms_pw(AEC *a, REAL d, REAL x_, float stepsize)
{
REAL e;
REAL ef;
a->x[a->j] = x_;
a->xf[a->j] = IIR1_highpass(a->Fx, x_); // pre-whitening of x
// calculate error value
// (mic signal - estimated mic signal from spk signal)
e = d;
if (a->hangover > 0) {
e -= a->dotp(a->w, a->x + a->j);
}
ef = IIR1_highpass(a->Fe, e); // pre-whitening of e
// optimize: iterative dotp(xf, xf)
a->dotp_xf_xf += (a->xf[a->j] * a->xf[a->j] - a->xf[a->j + NLMS_LEN - 1] * a->xf[a->j + NLMS_LEN - 1]);
if (stepsize > 0.0f) {
// calculate variable step size
REAL mikro_ef = stepsize * ef / a->dotp_xf_xf;
#ifdef DISABLE_ORC
// update tap weights (filter learning)
int i;
for (i = 0; i < NLMS_LEN; i += 2) {
// optimize: partial loop unrolling
a->w[i] += mikro_ef * a->xf[i + a->j];
a->w[i + 1] += mikro_ef * a->xf[i + a->j + 1];
}
#else
update_tap_weights(a->w, &a->xf[a->j], mikro_ef, NLMS_LEN);
#endif
}
if (--(a->j) < 0) {
// optimize: decrease number of memory copies
a->j = NLMS_EXT;
memmove(a->x + a->j + 1, a->x, (NLMS_LEN - 1) * sizeof(REAL));
memmove(a->xf + a->j + 1, a->xf, (NLMS_LEN - 1) * sizeof(REAL));
}
// Saturation
if (e > MAXPCM) {
return MAXPCM;
} else if (e < -MAXPCM) {
return -MAXPCM;
} else {
return e;
}
}
int AEC_doAEC(AEC *a, int d_, int x_)
{
REAL d = (REAL) d_;
REAL x = (REAL) x_;
// Mic Highpass Filter - to remove DC
d = IIR_HP_highpass(a->acMic, d);
// Mic Highpass Filter - cut-off below 300Hz
d = FIR_HP_300Hz_highpass(a->cutoff, d);
// Amplify, for e.g. Soundcards with -6dB max. volume
d *= a->gain;
// Spk Highpass Filter - to remove DC
x = IIR_HP_highpass(a->acSpk, x);
// Double Talk Detector
a->stepsize = AEC_dtd(a, d, x);
// Leaky (ageing of vector w)
AEC_leaky(a);
// Acoustic Echo Cancellation
d = AEC_nlms_pw(a, d, x, a->stepsize);
#if 0
if (fdwdisplay >= 0) {
if (++dumpcnt >= (WIDEB*RATE/10)) {
// wdisplay creates 10 dumps per seconds = large CPU load!
dumpcnt = 0;
write(fdwdisplay, ws, DUMP_LEN*sizeof(float));
// we don't check return value. This is not production quality!!!
memset(ws, 0, sizeof(ws));
} else {
int i;
for (i = 0; i < DUMP_LEN; i += 2) {
// optimize: partial loop unrolling
ws[i] += w[i];
ws[i + 1] += w[i + 1];
}
}
}
#endif
return (int) d;
}
|