-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate_tDCF_asvspoof19.py
126 lines (102 loc) · 5.27 KB
/
evaluate_tDCF_asvspoof19.py
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
import os
import numpy as np
import eval_metrics as em
import matplotlib.pyplot as plt
def compute_eer_and_tdcf(cm_score_file, path_to_database, visualize=False):
asv_score_file = os.path.join(path_to_database, 'LA/ASVspoof2019_LA_asv_scores/ASVspoof2019.LA.asv.eval.gi.trl.scores.txt')
# Fix tandem detection cost function (t-DCF) parameters
Pspoof = 0.05
cost_model = {
'Pspoof': Pspoof, # Prior probability of a spoofing attack
'Ptar': (1 - Pspoof) * 0.99, # Prior probability of target speaker
'Pnon': (1 - Pspoof) * 0.01, # Prior probability of nontarget speaker
'Cmiss_asv': 1, # Cost of ASV system falsely rejecting target speaker
'Cfa_asv': 10, # Cost of ASV system falsely accepting nontarget speaker
'Cmiss_cm': 1, # Cost of CM system falsely rejecting target speaker
'Cfa_cm': 10, # Cost of CM system falsely accepting spoof
}
# Load organizers' ASV scores
asv_data = np.genfromtxt(asv_score_file, dtype=str)
asv_sources = asv_data[:, 0]
asv_keys = asv_data[:, 1]
asv_scores = asv_data[:, 2].astype(np.float)
# Load CM scores
cm_data = np.genfromtxt(cm_score_file, dtype=str)
cm_utt_id = cm_data[:, 0]
cm_sources = cm_data[:, 1]
cm_keys = cm_data[:, 2]
cm_scores = cm_data[:, 3].astype(np.float)
# Extract target, nontarget, and spoof scores from the ASV scores
tar_asv = asv_scores[asv_keys == 'target']
non_asv = asv_scores[asv_keys == 'nontarget']
spoof_asv = asv_scores[asv_keys == 'spoof']
# Extract bona fide (real human) and spoof scores from the CM scores
bona_cm = cm_scores[cm_keys == 'bonafide']
spoof_cm = cm_scores[cm_keys == 'spoof']
# EERs of the standalone systems and fix ASV operating point to EER threshold
eer_asv, asv_threshold = em.compute_eer(tar_asv, non_asv)
eer_cm = em.compute_eer(bona_cm, spoof_cm)[0]
[Pfa_asv, Pmiss_asv, Pmiss_spoof_asv] = em.obtain_asv_error_rates(tar_asv, non_asv, spoof_asv, asv_threshold)
# Compute t-DCF
tDCF_curve, CM_thresholds = em.compute_tDCF(bona_cm, spoof_cm, Pfa_asv, Pmiss_asv, Pmiss_spoof_asv, cost_model)
# Minimum t-DCF
min_tDCF_index = np.argmin(tDCF_curve)
min_tDCF = tDCF_curve[min_tDCF_index]
# print('ASV SYSTEM')
# print(' EER = {:8.5f} % (Equal error rate (target vs. nontarget discrimination)'.format(eer_asv * 100))
# print(' Pfa = {:8.5f} % (False acceptance rate of nontargets)'.format(Pfa_asv * 100))
# print(' Pmiss = {:8.5f} % (False rejection rate of targets)'.format(Pmiss_asv * 100))
# print(' 1-Pmiss,spoof = {:8.5f} % (Spoof false acceptance rate)'.format((1 - Pmiss_spoof_asv) * 100))
print('\nCM SYSTEM')
print(' EER = {:8.5f} % (Equal error rate for countermeasure)'.format(eer_cm * 100))
print('\nTANDEM')
print(' min-tDCF = {:8.5f}'.format(min_tDCF))
if visualize:
# Visualize ASV scores and CM scores
plt.figure()
ax = plt.subplot(121)
plt.hist(tar_asv, histtype='step', density=True, bins=50, label='Target')
plt.hist(non_asv, histtype='step', density=True, bins=50, label='Nontarget')
plt.hist(spoof_asv, histtype='step', density=True, bins=50, label='Spoof')
plt.plot(asv_threshold, 0, 'o', markersize=10, mfc='none', mew=2, clip_on=False, label='EER threshold')
plt.legend()
plt.xlabel('ASV score')
plt.ylabel('Density')
plt.title('ASV score histogram')
ax = plt.subplot(122)
plt.hist(bona_cm, histtype='step', density=True, bins=50, label='Bona fide')
plt.hist(spoof_cm, histtype='step', density=True, bins=50, label='Spoof')
plt.legend()
plt.xlabel('CM score')
# plt.ylabel('Density')
plt.title('CM score histogram')
plt.savefig(cm_score_file[:-4]+'1.png')
# Plot t-DCF as function of the CM threshold.
plt.figure()
plt.plot(CM_thresholds, tDCF_curve)
plt.plot(CM_thresholds[min_tDCF_index], min_tDCF, 'o', markersize=10, mfc='none', mew=2)
plt.xlabel('CM threshold index (operating point)')
plt.ylabel('Norm t-DCF')
plt.title('Normalized tandem t-DCF')
plt.plot([np.min(CM_thresholds), np.max(CM_thresholds)], [1, 1], '--', color='black')
plt.legend(('t-DCF', 'min t-DCF ({:.5f})'.format(min_tDCF), 'Arbitrarily bad CM (Norm t-DCF=1)'))
plt.xlim([np.min(CM_thresholds), np.max(CM_thresholds)])
plt.ylim([0, 1.5])
plt.savefig(cm_score_file[:-4]+'2.png')
plt.show()
return eer_cm, min_tDCF
if __name__ == "__main__":
# Replace CM scores with your own scores or provide score file as the first argument.
cm_score_file = 'cm_score_ocnn.txt'
# Replace ASV scores with organizers' scores or provide score file as the second argument.
# path_to_database = '/home/yzh298/Downloads/DS_10283_3336/'
path_to_database = '/data/neil/DS_10283_3336/' # if run on GPU
# args = sys.argv
# if len(args) > 1:
# if len(args) != 3:
# print('USAGE: python evaluate_tDCF_asvspoof19.py <CM_SCOREFILE> <ASV_SCOREFILE>')
# exit()
# else:
# cm_score_file = args[1]
# asv_score_file = args[2]
compute_eer_and_tdcf(cm_score_file, path_to_database)