-
Notifications
You must be signed in to change notification settings - Fork 0
/
learn_svdtmapmacro.m
184 lines (150 loc) · 5.28 KB
/
learn_svdtmapmacro.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
173
174
175
176
177
178
179
180
181
182
183
184
clear all;
% Create the mapping table first by running
cd D:\Corpus\GMU\audiodata
load FileWav16kHz
% see create_gmm for what to be loaded by FileWav16kHz
doNormalize = 1; % Already normalized
root = 'wav16kHzMlfDur';
files = dir(fullfile(root, '*.mat'));
%% Class definition
% Gender
Class(1, 1).name = 'Male';
Class(1, 1).IsTrain = IsTrain.IsMale;
Class(1, 1).IsTest = IsTest.IsMale;
Class(1, 2).name = 'Female';
Class(1, 2).IsTrain = IsTrain.IsFemale;
Class(1, 2).IsTest = IsTest.IsFemale;
% Native
Class(2, 1).name = 'Native';
Class(2, 1).IsTrain = IsTrain.IsNative;
Class(2, 1).IsTest = IsTest.IsNative;
Class(2, 2).name = 'NonNative';
Class(2, 2).IsTrain = IsTrain.IsNonNative;
Class(2, 2).IsTest = IsTest.IsNonNative;
%% Read feature vectors and SVD training
Nsvd = 60;
for i = 1:2
for j = 1:2
clear PhVecs WdVecs PhVecsNosp WdVecsNosp
idx = find(Class(i, j).IsTrain);
for k = 1:length(idx)
V = load(fullfile(root, files(idx(k)).name));
PhVecs(:, k) = V.Phn.dur;
WdVecs(:, k) = V.Wrd.dur;
PhVecsNosp(:, k) = V.PhnNosp.dur;
WdVecsNosp(:, k) = V.WrdNosp.dur;
% Normalize?
PhVecs(:, k) = PhVecs(:, k)/sum(PhVecs(:, k));
WdVecs(:, k) = WdVecs(:, k)/sum(WdVecs(:, k));
PhVecsNosp(:, k) = PhVecsNosp(:, k)/sum(PhVecsNosp(:, k));
WdVecsNosp(:, k) = WdVecsNosp(:, k)/sum(WdVecsNosp(:, k));
end
CellPhn{i, j} = PhVecs;
CellWrd{i, j} = WdVecs;
CellPhnNosp{i, j} = PhVecsNosp;
CellWrdNosp{i, j} = WdVecsNosp;
% SVD train
[UP, SP, VP] = svd(CellPhn{i, j});
[UW, SW, VW] = svd(CellWrd{i, j});
[UPN, SP, VP] = svd(CellPhnNosp{i, j});
[UWN, SW, VW] = svd(CellWrdNosp{i, j});
SvdClass(i, j).UP = UP(:, 1:Nsvd);
SvdClass(i, j).UW = UW(:, 1:Nsvd);
SvdClass(i, j).UPN = UPN(:, 1:Nsvd);
SvdClass(i, j).UWN = UWN(:, 1:Nsvd);
end
end
%% Performance Evaluation SVD - Phone Representation
NsvdTest = 60;
for i = 1 : 2
for j = 1 : 2
Fidx = find(Class(i, j).IsTest);
U1 = SvdClass(i, 1).UP(:, 1:NsvdTest);
U1_ = inv(U1'*U1)*U1';
U2 = SvdClass(i, 2).UP(:, 1:NsvdTest);
U2_ = inv(U2'*U2)*U2';
clear Err1 Err2
for k = 1 : length(Fidx)
Z = load(fullfile(root, files(Fidx(k)).name));
Coef1 = U1_*Z.Phn.dur;
VecHat1 = U1*Coef1;
Err1(k).n2 = norm(VecHat1 - Z.Phn.dur);
Coef2 = U2_*Z.Phn.dur;
VecHat2 = U2*Coef2;
Err2(k).n2 = norm(VecHat2 - Z.Phn.dur);
end
PC(i, j) = mean([Err1.n2] < [Err2.n2]);
end
PC(i, j) = 1 - PC(i, j);
end
PC
%% Performance Evaluation SVD - Phone Representation - No space
NsvdTest = 60;
for i = 1 : 2
for j = 1 : 2
Fidx = find(Class(i, j).IsTest);
U1 = SvdClass(i, 1).UPN(:, 1:NsvdTest);
U1_ = inv(U1'*U1)*U1';
U2 = SvdClass(i, 2).UPN(:, 1:NsvdTest);
U2_ = inv(U2'*U2)*U2';
clear Err1 Err2
for k = 1 : length(Fidx)
Z = load(fullfile(root, files(Fidx(k)).name));
Coef1 = U1_*Z.PhnNosp.dur;
VecHat1 = U1*Coef1;
Err1(k).n2 = norm(VecHat1 - Z.PhnNosp.dur);
Coef2 = U2_*Z.PhnNosp.dur;
VecHat2 = U2*Coef2;
Err2(k).n2 = norm(VecHat2 - Z.PhnNosp.dur);
end
PC(i, j) = mean([Err1.n2] < [Err2.n2]);
end
PC(i, j) = 1 - PC(i, j);
end
PC
%% Performance Evaluation SVD - Word Representation
for i = 1 : 2
for j = 1 : 2
Fidx = find(Class(i, j).IsTest);
U1 = SvdClass(i, 1).UW(:, 1:NsvdTest);
U1_ = inv(U1'*U1)*U1';
U2 = SvdClass(i, 2).UW(:, 1:NsvdTest);
U2_ = inv(U2'*U2)*U2';
clear Err1 Err2
for k = 1 : length(Fidx)
Z = load(fullfile(root, files(Fidx(k)).name));
Coef1 = U1_*Z.Wrd.dur;
VecHat1 = U1*Coef1;
Err1(k).n2 = norm(VecHat1 - Z.Wrd.dur);
Coef2 = U2_*Z.Wrd.dur;
VecHat2 = U2*Coef2;
Err2(k).n2 = norm(VecHat2 - Z.Wrd.dur);
end
PCW(i, j) = mean([Err1.n2] < [Err2.n2]);
end
PCW(i, j) = 1 - PCW(i, j);
end
PCW
%% Performance Evaluation SVD - Word Representation - No space
for i = 1 : 2
for j = 1 : 2
Fidx = find(Class(i, j).IsTest);
U1 = SvdClass(i, 1).UWN(:, 1:NsvdTest);
U1_ = inv(U1'*U1)*U1';
U2 = SvdClass(i, 2).UWN(:, 1:NsvdTest);
U2_ = inv(U2'*U2)*U2';
clear Err1 Err2
for k = 1 : length(Fidx)
Z = load(fullfile(root, files(Fidx(k)).name));
Coef1 = U1_*Z.WrdNosp.dur;
VecHat1 = U1*Coef1;
Err1(k).n2 = norm(VecHat1 - Z.WrdNosp.dur);
Coef2 = U2_*Z.WrdNosp.dur;
VecHat2 = U2*Coef2;
Err2(k).n2 = norm(VecHat2 - Z.WrdNosp.dur);
end
PCW(i, j) = mean([Err1.n2] < [Err2.n2]);
end
PCW(i, j) = 1 - PCW(i, j);
end
PCW