-
Notifications
You must be signed in to change notification settings - Fork 25
/
saf_kalman.m
172 lines (154 loc) · 6.2 KB
/
saf_kalman.m
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
function out = saf_kalman(mic, spk, frame_size)
out_len = min(length(mic),length(spk));
out = zeros(out_len,1);
out_num = floor(out_len/frame_size);
st = init(frame_size);
for i = 1 : out_num
mic_frame = mic((i - 1) * frame_size + 1 : i * frame_size);
spk_frame = spk((i - 1) * frame_size + 1 : i * frame_size);
[st, out_frame] = saf_process(st, mic_frame, spk_frame);
out(1+(i-1)*frame_size:i*frame_size) = out_frame;
end
function st = init(frame_size)
% static para
st.frame_len = frame_size;
st.K = frame_size*2;
st.half_bin = st.K / 2 + 1;
st.win_len = st.K * 2;
st.notch_radius = .982;
st.notch_mem = zeros(2,1);
st.memX = 0;
st.memD = 0;
st.memE = 0;
st.preemph = .98;
% lp win
win_st = load('win_para_update.mat');
st.win_global = win_st.win;
% subband para
st.ana_win_echo = zeros(1, st.win_len);
st.ana_win_far = zeros(1, st.win_len);
st.sys_win = zeros(1, st.win_len);
st.tap = 15;
st.subband_in = zeros(st.half_bin, st.tap);
st.subband_adf = zeros(st.half_bin, st.tap);
%kalman para
st.Ryu = ones(st.half_bin,st.tap,st.tap)*5;
st.w_cov = ones(st.half_bin, 1)*0.1;
st.v_cov = ones(st.half_bin, 1)*0.001;
st.gain = zeros(st.half_bin,st.tap);
% nlp
st.Eh = zeros(st.half_bin,1);
st.Yh = zeros(st.half_bin,1);
st.est_ps = zeros(st.half_bin,1);
st.spec_ave = 0.01;
st.Pey = 0;
st.Pyy = 0;
st.beta0 = 0.016;
st.beta_max = st.beta0/4;
st.min_leak = 0.005;
st.echo_noise_ps = 0;
st.adapt_cnt=0;
st.res_old_ps = 0;
st.suppress_gain = 60;
st.wiener_gain = zeros(st.half_bin,1);
st.gain_floor = ones(st.half_bin,1).*0.01;
end
function [out,mem] = filter_dc_notch16(in, radius, len, mem)
out = zeros(size(in));
den2 = radius*radius + .7*(1-radius)*(1-radius);
for ii=1:len
vin = in(ii);
vout = mem(1) + vin;
mem(1) = mem(2) + 2*(-vin + radius*vout);
mem(2) = vin - (den2*vout);
out(ii) = radius*vout;
end
end
function [st, out] = saf_process(st, mic_frame, spk_frame)
N = st.frame_len;
K = st.K;
[mic_in, st.notch_mem] = filter_dc_notch16(mic_frame, st.notch_radius, N, st.notch_mem);
st.ana_win_echo = [st.ana_win_echo(N+1:end), mic_in'];
ana_win_echo_windowed = st.win_global .* st.ana_win_echo;
ana_wined_echo = ana_win_echo_windowed(1:K)+ana_win_echo_windowed(K+1:2*K);
fft_out_echo = fft(ana_wined_echo);
st.ana_win_far = [st.ana_win_far(N + 1:end), spk_frame'];
ana_win_far_windowed = st.win_global .* st.ana_win_far;
ana_wined_far = ana_win_far_windowed(1:K)+ana_win_far_windowed(K+1:2*K);
fft_out_far = fft(ana_wined_far, K);
st.subband_in = [fft_out_far(1:st.half_bin)', st.subband_in(:,1:st.tap-1)];
subband_adf_out = sum(st.subband_adf .* st.subband_in,2);
subband_adf_err = fft_out_echo(1:st.half_bin)' - subband_adf_out;
% kalman update
for j = 1:st.half_bin
%update sigmal v
st.v_cov(j) = 0.99*st.v_cov(j) + 0.01*(subband_adf_err(j)*subband_adf_err(j)');
Rmu = squeeze(st.Ryu(j,:,:)) + eye(st.tap).*st.w_cov(j);
Re = real(st.subband_in(j,:) * Rmu * st.subband_in(j,:)') + st.v_cov(j);
gain = (Rmu * st.subband_in(j,:)') ./ (Re+1e-10);
phi = gain .* subband_adf_err(j);
st.subband_adf(j,:) = st.subband_adf(j,:) + phi.';
st.Ryu(j,:,:) = (eye(st.tap) - gain * st.subband_in(j,:)) * Rmu;
%update sigmal w
st.w_cov(j) = 0.99* st.w_cov(j) + 0.01 * (sqrt(phi' * phi)/st.tap);
end
% compose subband
if(1)
% nlp
[st,nlpout] = nlpProcess(st,subband_adf_err, subband_adf_out);
ifft_in = [nlpout', fliplr(conj(nlpout(2:end-1)'))];
else
ifft_in = [subband_adf_err', fliplr(conj(subband_adf_err(2:end-1)'))];
end
fft_out = real(ifft(ifft_in));
win_in = [fft_out, fft_out];
comp_out = win_in .* st.win_global;
st.sys_win = st.sys_win + comp_out;
out = st.sys_win(1 : N);
st.sys_win = [st.sys_win(N + 1 : end), zeros(1, N)];
st.adapt_cnt = st.adapt_cnt + 1;
end
function [st,nlp_out] = nlpProcess(st, error, est_echo)
st.est_ps = abs(est_echo).^2;
res_ps = abs(error).^2;
Eh_curr = res_ps - st.Eh;
Yh_curr = st.est_ps - st.Yh;
Pey = sum(Eh_curr.*Yh_curr);
Pyy = sum(Yh_curr.*Yh_curr);
st.Eh = (1-st.spec_ave)*st.Eh + st.spec_ave*(res_ps);
st.Yh = (1-st.spec_ave)*st.Yh + st.spec_ave*(st.est_ps);
Syy = sum(st.est_ps);
See = sum(res_ps);
Pyy = sqrt(Pyy);
Pey = Pey/(Pyy+1e-10);
tmp32 = st.beta0*Syy/See;
alpha = min(tmp32, st.beta_max);
st.Pyy = (1-alpha)*st.Pyy + alpha*Pyy;
st.Pey = (1-alpha)*st.Pey + alpha*Pey;
st.Pyy = max(st.Pyy,1);
st.Pey = max(st.Pey, st.Pyy*st.min_leak);
st.Pey = min(st.Pey, st.Pyy);
leak = st.Pey/st.Pyy;
if(leak>0.5)
leak=1;
end
residual_ps = leak*st.est_ps*st.suppress_gain;
if(st.adapt_cnt==0)
st.echo_noise_ps = residual_ps;
else
st.echo_noise_ps = max(0.85*st.echo_noise_ps, residual_ps);
end
st.echo_noise_ps = max(st.echo_noise_ps, 1e-10);
postser = res_ps./st.echo_noise_ps -1;
postser = min(postser,100);
if(st.adapt_cnt==0)
st.res_old_ps = res_ps;
end
prioriser = 0.5.*max(0, postser)+0.5.*(st.res_old_ps./st.echo_noise_ps);
prioriser = min(prioriser,100);
st.wiener_gain = prioriser./(prioriser+1);
st.wiener_gain(1:100) = max(st.wiener_gain(1:100),0.1);
st.res_old_ps = 0.8*st.res_old_ps + 0.2*st.wiener_gain.*res_ps;
nlp_out = st.wiener_gain.*error;
end
end