|
2
|
1 /* aec.h
|
|
|
2 *
|
|
|
3 * Copyright (C) DFS Deutsche Flugsicherung (2004, 2005).
|
|
|
4 * All Rights Reserved.
|
|
|
5 * Author: Andre Adrian
|
|
|
6 *
|
|
|
7 * Acoustic Echo Cancellation Leaky NLMS-pw algorithm
|
|
|
8 *
|
|
|
9 * Version 0.3 filter created with www.dsptutor.freeuk.com
|
|
|
10 * Version 0.3.1 Allow change of stability parameter delta
|
|
|
11 * Version 0.4 Leaky Normalized LMS - pre whitening algorithm
|
|
|
12 */
|
|
|
13
|
|
|
14 #ifndef _AEC_H /* include only once */
|
|
|
15
|
|
|
16 // use double if your CPU does software-emulation of float
|
|
|
17 typedef float REAL;
|
|
|
18
|
|
|
19 /* dB Values */
|
|
|
20 const REAL M0dB = 1.0f;
|
|
|
21 const REAL M3dB = 0.71f;
|
|
|
22 const REAL M6dB = 0.50f;
|
|
|
23 const REAL M9dB = 0.35f;
|
|
|
24 const REAL M12dB = 0.25f;
|
|
|
25 const REAL M18dB = 0.125f;
|
|
|
26 const REAL M24dB = 0.063f;
|
|
|
27
|
|
|
28 /* dB values for 16bit PCM */
|
|
|
29 /* MxdB_PCM = 32767 * 10 ^(x / 20) */
|
|
|
30 const REAL M10dB_PCM = 10362.0f;
|
|
|
31 const REAL M20dB_PCM = 3277.0f;
|
|
|
32 const REAL M25dB_PCM = 1843.0f;
|
|
|
33 const REAL M30dB_PCM = 1026.0f;
|
|
|
34 const REAL M35dB_PCM = 583.0f;
|
|
|
35 const REAL M40dB_PCM = 328.0f;
|
|
|
36 const REAL M45dB_PCM = 184.0f;
|
|
|
37 const REAL M50dB_PCM = 104.0f;
|
|
|
38 const REAL M55dB_PCM = 58.0f;
|
|
|
39 const REAL M60dB_PCM = 33.0f;
|
|
|
40 const REAL M65dB_PCM = 18.0f;
|
|
|
41 const REAL M70dB_PCM = 10.0f;
|
|
|
42 const REAL M75dB_PCM = 6.0f;
|
|
|
43 const REAL M80dB_PCM = 3.0f;
|
|
|
44 const REAL M85dB_PCM = 2.0f;
|
|
|
45 const REAL M90dB_PCM = 1.0f;
|
|
|
46
|
|
|
47 const REAL MAXPCM = 32767.0f;
|
|
|
48
|
|
|
49 /* Design constants (Change to fine tune the algorithms */
|
|
|
50
|
|
|
51 /* The following values are for hardware AEC and studio quality
|
|
|
52 * microphone */
|
|
|
53
|
|
|
54 /* NLMS filter length in taps (samples). A longer filter length gives
|
|
|
55 * better Echo Cancellation, but maybe slower convergence speed and
|
|
|
56 * needs more CPU power (Order of NLMS is linear) */
|
|
|
57 #define NLMS_LEN (100*WIDEB*8)
|
|
|
58
|
|
|
59 /* Vector w visualization length in taps (samples).
|
|
|
60 * Must match argv value for wdisplay.tcl */
|
|
|
61 #define DUMP_LEN (40*WIDEB*8)
|
|
|
62
|
|
|
63 /* minimum energy in xf. Range: M70dB_PCM to M50dB_PCM. Should be equal
|
|
|
64 * to microphone ambient Noise level */
|
|
|
65 const REAL NoiseFloor = M55dB_PCM;
|
|
|
66
|
|
|
67 /* Leaky hangover in taps.
|
|
|
68 */
|
|
|
69 const int Thold = 60 * WIDEB * 8;
|
|
|
70
|
|
|
71 // Adrian soft decision DTD
|
|
|
72 // left point. X is ratio, Y is stepsize
|
|
|
73 const float STEPX1 = 1.0, STEPY1 = 1.0;
|
|
|
74 // right point. STEPX2=2.0 is good double talk, 3.0 is good single talk.
|
|
|
75 const float STEPX2 = 2.5, STEPY2 = 0;
|
|
|
76 const float ALPHAFAST = 1.0f / 100.0f;
|
|
|
77 const float ALPHASLOW = 1.0f / 20000.0f;
|
|
|
78
|
|
|
79
|
|
|
80
|
|
|
81 /* Ageing multiplier for LMS memory vector w */
|
|
|
82 const REAL Leaky = 0.9999f;
|
|
|
83
|
|
|
84 /* Double Talk Detector Speaker/Microphone Threshold. Range <=1
|
|
|
85 * Large value (M0dB) is good for Single-Talk Echo cancellation,
|
|
|
86 * small value (M12dB) is good for Doulbe-Talk AEC */
|
|
|
87 const REAL GeigelThreshold = M6dB;
|
|
|
88
|
|
|
89 /* for Non Linear Processor. Range >0 to 1. Large value (M0dB) is good
|
|
|
90 * for Double-Talk, small value (M12dB) is good for Single-Talk */
|
|
|
91 const REAL NLPAttenuation = M12dB;
|
|
|
92
|
|
|
93 /* Below this line there are no more design constants */
|
|
|
94
|
|
|
95
|
|
|
96 /* Exponential Smoothing or IIR Infinite Impulse Response Filter */
|
|
|
97 class IIR_HP {
|
|
|
98 REAL x;
|
|
|
99
|
|
|
100 public:
|
|
|
101 IIR_HP() {
|
|
|
102 x = 0.0f;
|
|
|
103 }
|
|
|
104
|
|
|
105 REAL highpass(REAL in) {
|
|
|
106 const REAL a0 = 0.01f; /* controls Transfer Frequency */
|
|
|
107 /* Highpass = Signal - Lowpass. Lowpass = Exponential Smoothing */
|
|
|
108 x += a0 * (in - x);
|
|
|
109 return in - x;
|
|
|
110 };
|
|
|
111 };
|
|
|
112
|
|
|
113 #if WIDEB==1
|
|
|
114 /* 17 taps FIR Finite Impulse Response filter
|
|
|
115 * Coefficients calculated with
|
|
|
116 * www.dsptutor.freeuk.com/KaiserFilterDesign/KaiserFilterDesign.html
|
|
|
117 */
|
|
|
118 class FIR_HP_300Hz {
|
|
|
119 REAL z[18];
|
|
|
120
|
|
|
121 public:
|
|
|
122 FIR_HP_300Hz() {
|
|
|
123 memset(this, 0, sizeof(FIR_HP_300Hz));
|
|
|
124 }
|
|
|
125
|
|
|
126 REAL highpass(REAL in) {
|
|
|
127 const REAL a[18] = {
|
|
|
128 // Kaiser Window FIR Filter, Filter type: High pass
|
|
|
129 // Passband: 300.0 - 4000.0 Hz, Order: 16
|
|
|
130 // Transition band: 75.0 Hz, Stopband attenuation: 10.0 dB
|
|
|
131 -0.034870606, -0.039650206, -0.044063766, -0.04800318,
|
|
|
132 -0.051370874, -0.054082647, -0.056070227, -0.057283327,
|
|
|
133 0.8214126, -0.057283327, -0.056070227, -0.054082647,
|
|
|
134 -0.051370874, -0.04800318, -0.044063766, -0.039650206,
|
|
|
135 -0.034870606, 0.0
|
|
|
136 };
|
|
|
137 memmove(z + 1, z, 17 * sizeof(REAL));
|
|
|
138 z[0] = in;
|
|
|
139 REAL sum0 = 0.0, sum1 = 0.0;
|
|
|
140 int j;
|
|
|
141
|
|
|
142 for (j = 0; j < 18; j += 2) {
|
|
|
143 // optimize: partial loop unrolling
|
|
|
144 sum0 += a[j] * z[j];
|
|
|
145 sum1 += a[j + 1] * z[j + 1];
|
|
|
146 }
|
|
|
147 return sum0 + sum1;
|
|
|
148 }
|
|
|
149 };
|
|
|
150
|
|
|
151 #else
|
|
|
152
|
|
|
153 /* 35 taps FIR Finite Impulse Response filter
|
|
|
154 * Passband 150Hz to 4kHz for 8kHz sample rate, 300Hz to 8kHz for 16kHz
|
|
|
155 * sample rate.
|
|
|
156 * Coefficients calculated with
|
|
|
157 * www.dsptutor.freeuk.com/KaiserFilterDesign/KaiserFilterDesign.html
|
|
|
158 */
|
|
|
159 class FIR_HP_300Hz {
|
|
|
160 REAL z[36];
|
|
|
161
|
|
|
162 public:
|
|
|
163 FIR_HP_300Hz() {
|
|
|
164 memset(this, 0, sizeof(FIR_HP_300Hz));
|
|
|
165 }
|
|
|
166
|
|
|
167 REAL highpass(REAL in) {
|
|
|
168 const REAL a[36] = {
|
|
|
169 // Kaiser Window FIR Filter, Filter type: High pass
|
|
|
170 // Passband: 150.0 - 4000.0 Hz, Order: 34
|
|
|
171 // Transition band: 34.0 Hz, Stopband attenuation: 10.0 dB
|
|
|
172 -0.016165324, -0.017454365, -0.01871232, -0.019931411,
|
|
|
173 -0.021104068, -0.022222936, -0.02328091, -0.024271343,
|
|
|
174 -0.025187887, -0.02602462, -0.026776174, -0.027437767,
|
|
|
175 -0.028004972, -0.028474221, -0.028842418, -0.029107114,
|
|
|
176 -0.02926664, 0.8524841, -0.02926664, -0.029107114,
|
|
|
177 -0.028842418, -0.028474221, -0.028004972, -0.027437767,
|
|
|
178 -0.026776174, -0.02602462, -0.025187887, -0.024271343,
|
|
|
179 -0.02328091, -0.022222936, -0.021104068, -0.019931411,
|
|
|
180 -0.01871232, -0.017454365, -0.016165324, 0.0
|
|
|
181 };
|
|
|
182 memmove(z + 1, z, 35 * sizeof(REAL));
|
|
|
183 z[0] = in;
|
|
|
184 REAL sum0 = 0.0, sum1 = 0.0;
|
|
|
185 int j;
|
|
|
186
|
|
|
187 for (j = 0; j < 36; j += 2) {
|
|
|
188 // optimize: partial loop unrolling
|
|
|
189 sum0 += a[j] * z[j];
|
|
|
190 sum1 += a[j + 1] * z[j + 1];
|
|
|
191 }
|
|
|
192 return sum0 + sum1;
|
|
|
193 }
|
|
|
194 };
|
|
|
195 #endif
|
|
|
196
|
|
|
197 /* Recursive single pole IIR Infinite Impulse response High-pass filter
|
|
|
198 *
|
|
|
199 * Reference: The Scientist and Engineer's Guide to Digital Processing
|
|
|
200 *
|
|
|
201 * output[N] = A0 * input[N] + A1 * input[N-1] + B1 * output[N-1]
|
|
|
202 *
|
|
|
203 * X = exp(-2.0 * pi * Fc)
|
|
|
204 * A0 = (1 + X) / 2
|
|
|
205 * A1 = -(1 + X) / 2
|
|
|
206 * B1 = X
|
|
|
207 * Fc = cutoff freq / sample rate
|
|
|
208 */
|
|
|
209 class IIR1 {
|
|
|
210 REAL in0, out0;
|
|
|
211 REAL a0, a1, b1;
|
|
|
212
|
|
|
213 public:
|
|
|
214 IIR1() {
|
|
|
215 memset(this, 0, sizeof(IIR1));
|
|
|
216 }
|
|
|
217
|
|
|
218 void init(REAL Fc) {
|
|
|
219 b1 = expf(-2.0f * M_PI * Fc);
|
|
|
220 a0 = (1.0f + b1) / 2.0f;
|
|
|
221 a1 = -a0;
|
|
|
222 in0 = 0.0f;
|
|
|
223 out0 = 0.0f;
|
|
|
224 }
|
|
|
225
|
|
|
226 REAL highpass(REAL in) {
|
|
|
227 REAL out = a0 * in + a1 * in0 + b1 * out0;
|
|
|
228 in0 = in;
|
|
|
229 out0 = out;
|
|
|
230 return out;
|
|
|
231 }
|
|
|
232 };
|
|
|
233
|
|
|
234
|
|
|
235 /* Recursive two pole IIR Infinite Impulse Response filter
|
|
|
236 * Coefficients calculated with
|
|
|
237 * http://www.dsptutor.freeuk.com/IIRFilterDesign/IIRFiltDes102.html
|
|
|
238 */
|
|
|
239 class IIR2 {
|
|
|
240 REAL x[2], y[2];
|
|
|
241
|
|
|
242 public:
|
|
|
243 IIR2() {
|
|
|
244 memset(this, 0, sizeof(IIR2));
|
|
|
245 }
|
|
|
246
|
|
|
247 REAL highpass(REAL in) {
|
|
|
248 // Butterworth IIR filter, Filter type: HP
|
|
|
249 // Passband: 2000 - 4000.0 Hz, Order: 2
|
|
|
250 const REAL a[] = { 0.29289323f, -0.58578646f, 0.29289323f };
|
|
|
251 const REAL b[] = { 1.3007072E-16f, 0.17157288f };
|
|
|
252 REAL out =
|
|
|
253 a[0] * in + a[1] * x[0] + a[2] * x[1] - b[0] * y[0] - b[1] * y[1];
|
|
|
254
|
|
|
255 x[1] = x[0];
|
|
|
256 x[0] = in;
|
|
|
257 y[1] = y[0];
|
|
|
258 y[0] = out;
|
|
|
259 return out;
|
|
|
260 }
|
|
|
261 };
|
|
|
262
|
|
|
263
|
|
|
264 // Extention in taps to reduce mem copies
|
|
|
265 #define NLMS_EXT (10*8)
|
|
|
266
|
|
|
267 // block size in taps to optimize DTD calculation
|
|
|
268 #define DTD_LEN 16
|
|
|
269
|
|
|
270
|
|
|
271 class AEC {
|
|
|
272 // Time domain Filters
|
|
|
273 IIR_HP acMic, acSpk; // DC-level remove Highpass)
|
|
|
274 FIR_HP_300Hz cutoff; // 150Hz cut-off Highpass
|
|
|
275 REAL gain; // Mic signal amplify
|
|
|
276 IIR1 Fx, Fe; // pre-whitening Highpass for x, e
|
|
|
277
|
|
|
278 // Adrian soft decision DTD (Double Talk Detector)
|
|
|
279 REAL dfast, xfast;
|
|
|
280 REAL dslow, xslow;
|
|
|
281
|
|
|
282 // NLMS-pw
|
|
|
283 REAL x[NLMS_LEN + NLMS_EXT]; // tap delayed loudspeaker signal
|
|
|
284 REAL xf[NLMS_LEN + NLMS_EXT]; // pre-whitening tap delayed signal
|
|
|
285 REAL w[NLMS_LEN]; // tap weights
|
|
|
286 int j; // optimize: less memory copies
|
|
|
287 double dotp_xf_xf; // double to avoid loss of precision
|
|
|
288 float delta; // noise floor to stabilize NLMS
|
|
|
289
|
|
|
290 // AES
|
|
|
291 float aes_y2; // not in use!
|
|
|
292
|
|
|
293 // w vector visualization
|
|
|
294 REAL ws[DUMP_LEN]; // tap weights sums
|
|
|
295 int fdwdisplay; // TCP file descriptor
|
|
|
296 int dumpcnt; // wdisplay output counter
|
|
|
297
|
|
|
298 /* Double-Talk Detector
|
|
|
299 *
|
|
|
300 * in d: microphone sample (PCM as REALing point value)
|
|
|
301 * in x: loudspeaker sample (PCM as REALing point value)
|
|
|
302 * return: from 0 for doubletalk to 1.0 for single talk
|
|
|
303 */
|
|
|
304 float dtd(REAL d, REAL x);
|
|
|
305
|
|
3
|
306 void leaky();
|
|
2
|
307
|
|
|
308 /* Normalized Least Mean Square Algorithm pre-whitening (NLMS-pw)
|
|
|
309 * The LMS algorithm was developed by Bernard Widrow
|
|
|
310 * book: Haykin, Adaptive Filter Theory, 4. edition, Prentice Hall, 2002
|
|
|
311 *
|
|
|
312 * in d: microphone sample (16bit PCM value)
|
|
|
313 * in x_: loudspeaker sample (16bit PCM value)
|
|
|
314 * in stepsize: NLMS adaptation variable
|
|
|
315 * return: echo cancelled microphone sample
|
|
|
316 */
|
|
|
317 REAL nlms_pw(REAL d, REAL x_, float stepsize);
|
|
|
318
|
|
|
319 public:
|
|
|
320 // variables are public for visualization
|
|
|
321 int hangover;
|
|
|
322 float stepsize;
|
|
|
323 AEC();
|
|
|
324
|
|
|
325 /* Acoustic Echo Cancellation and Suppression of one sample
|
|
|
326 * in d: microphone signal with echo
|
|
|
327 * in x: loudspeaker signal
|
|
|
328 * return: echo cancelled microphone signal
|
|
|
329 */
|
|
3
|
330 int doAEC(int d_, int x_);
|
|
2
|
331
|
|
3
|
332 float getambient() {
|
|
2
|
333 return dfast;
|
|
|
334 };
|
|
3
|
335 void setambient(float Min_xf) {
|
|
2
|
336 dotp_xf_xf -= delta; // subtract old delta
|
|
|
337 delta = (NLMS_LEN-1) * Min_xf * Min_xf;
|
|
|
338 dotp_xf_xf += delta; // add new delta
|
|
|
339 };
|
|
3
|
340 void setgain(float gain_) {
|
|
2
|
341 gain = gain_;
|
|
|
342 };
|
|
3
|
343 void openwdisplay();
|
|
|
344 void setaes(float aes_y2_) {
|
|
2
|
345 aes_y2 = aes_y2_;
|
|
|
346 };
|
|
3
|
347 double max_dotp_xf_xf(double u);
|
|
2
|
348 };
|
|
|
349
|
|
|
350 #define _AEC_H
|
|
|
351 #endif
|
