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
288
289
290
291
292
|
#!/usr/bin/env python3
import sys
import math
from cmusphinx import lattice
import sphinxbase
def baseword(sym):
paren = sym.rfind('(')
if paren != -1:
return sym[0:paren]
else:
return sym
def load_lexicon(lexfile):
# load dict or filler file to get word to phone information
f = open(lexfile, 'r')
lines = f.readlines()
f.close()
lex = {}
for line in lines:
temp = line.split()
lex[temp[0]] = temp[1:len(temp)]
return lex
def load_denlat(latfile):
# Annotate the lattice with posterior probabilities
dag = lattice.Dag(latfile)
dag.remove_unreachable()
arc = []
for w in dag.nodes:
for l in w.exits:
if (w.entry, w.sym, l.dest.entry, l.dest.sym) not in arc:
arc.append((w.entry, w.sym, l.dest.entry, l.dest.sym))
lat = {}
for n in arc:
if n[1] != '<s>':
left = []
right = []
for m in arc:
if n[3] == '</s>' and (n[2], n[3], n[2]) not in right:
right.append((n[2], n[3], n[2]))
elif m[0] == n[2] and m[1] == n[3]:
if (m[0], m[1], m[2]) not in right:
right.append((m[0], m[1], m[2]))
if m[2] == n[0] and m[3] == n[1]:
if (m[0], m[1], m[2]) not in left:
left.append((m[0], m[1], m[2]))
if (n[0], n[1], n[2]) not in lat:
lat[(n[0], n[1], n[2])] = (left, right)
else:
for l in left:
if l not in lat[(n[0], n[1], n[2])][0]:
lat[(n[0], n[1], n[2])][0].append(l)
for r in right:
if r not in lat[(n[0], n[1], n[2])][1]:
lat[(n[0], n[1], n[2])][1].append(r)
return lat
def load_numlat(alignfile):
# read force alignment result which is numerator lattice
f = open(alignfile, 'r')
f.readline()
lines = f.readlines()
f.close()
wordseg = []
for line in lines:
temp = line.split()
if len(temp) == 4:
if '<s>' in temp:
wordseg.append((0, 1, '<s>'))
wordseg.append((1, int(temp[1]) + 2, '<sil>'))
elif '</s>' in temp:
wordseg.append((int(temp[0]) + 1, int(temp[1]) + 1, '<sil>'))
wordseg.append((int(temp[1]) + 1, int(temp[1]) + 1, '</s>'))
else:
wordseg.append((int(temp[0]) + 1, int(temp[1]) + 2, temp[3]))
if wordseg[1][2] == wordseg[2][2] and wordseg[1][2] == '<sil>' and wordseg[
1][1] == wordseg[2][0]:
sf = wordseg[1][0]
ef = wordseg[2][1]
wordseg.pop(1)
wordseg.pop(1)
wordseg.insert(1, (sf, ef, '<sil>'))
n = len(wordseg)
if wordseg[n - 2][2] == wordseg[n - 3][2] and wordseg[
n - 2][2] == '<sil>' and wordseg[n - 3][1] == wordseg[n - 2][0]:
sf = wordseg[n - 3][0]
ef = wordseg[n - 2][1]
x = wordseg.pop(n - 2)
y = wordseg.pop(n - 3)
wordseg.insert(n - 3, (sf, ef, '<sil>'))
numlat = {}
for i in range(1, len(wordseg) - 1):
sf = wordseg[i][0]
ef = wordseg[i][1]
word = wordseg[i][2]
left = []
right = []
left.append((wordseg[i - 1][0], wordseg[i - 1][2], wordseg[i - 1][1]))
right.append((wordseg[i + 1][0], wordseg[i + 1][2], wordseg[i + 1][1]))
numlat[(sf, word, ef)] = (left, right)
return numlat
def conv_lat_format(key, lat, filler, lm):
# add word segment IDs and
# use word IDs to present previsou and next word segments
# also assign unigram language model score to each word segment
newlat = []
i = 0
for sf, word, ef in key:
i = i + 1
if word in filler:
score = math.log(0.1)
else:
score = lm.prob([baseword(word)])
left = []
for context in lat[(sf, word, ef)][0]:
if '<s>' in context:
ID = 0
else:
ID = key.index(context) + 1
left.append(ID)
right = []
for context in lat[(sf, word, ef)][1]:
if '</s>' in context:
ID = 0
else:
ID = key.index(context) + 1
right.append(ID)
newlat.append((i, word, sf, ef, score, left, right))
return newlat
def add_numlat_into_denlat(numlat, denlat):
# add numerator lattice into denominator lattice
n_arc = len(denlat)
for i, word, sf, ef, score, left, right in numlat:
aid = i + n_arc
lids = []
for ids in left:
if ids == 0:
lids.append(ids)
else:
lids.append(ids + n_arc)
rids = []
for ids in right:
if ids == 0:
rids.append(ids)
else:
rids.append(ids + n_arc)
denlat.append((aid, word, sf, ef, score, lids, rids))
return denlat
def output_lattice(lat, n_true_arc, outlatfile):
# write lattice to a file
f = open(outlatfile, 'w')
f.write('Total arcs:\n')
f.write('%d\n' % len(lat))
f.write('True arcs:\n')
f.write('%d\n' % n_true_arc)
f.write(
'arc_id, arc_name, start frame, end frame, lmscore, number of preceding acrs, number of succeeding arcs, preceding arc_ids, succeeding arc_ids\n'
)
for ID, word, sf, ef, score, left, right in lat:
if score is None:
break
line = str(ID) + ' ' + word + ' ' + str(sf) + ' ' + str(
ef) + ' ' + str(score) + ' ' + str(len(left)) + ' ' + str(
len(right)) + ' <'
for lid in left:
line = line + ' ' + str(lid)
line = line + ' >'
for rid in right:
line = line + ' ' + str(rid)
f.write("%s\n" % line)
f.close()
def write_lat(lm, filler, filelist, filecount, fileoffset, denlatdir,
numlatdir, outdir):
# write lattice for mmi training
f = open(filelist, 'r')
files = f.readlines()
f.close()
start = int(fileoffset)
end = int(fileoffset) + int(filecount)
if end > len(files):
end = len(files)
for i in range(start, end):
line = files[i]
fileid = line.strip('\n')
alignfile = "%s/%s.wdseg" % (numlatdir, fileid)
latfile = "%s/%s.lat.gz" % (denlatdir, fileid)
print("process sent: %s" % fileid)
print("\t load numerator lattice %s ... " % alignfile)
try:
numlat = load_numlat(alignfile)
except IOError:
print("\t can't open numerator lattice %s to read" % alignfile)
continue
print("\t load denominator lattice %s ..." % latfile)
try:
den_wordseg = load_denlat(latfile)
except IOError:
print("\t can't open denominator lattice %s to read" % latfile)
continue
print("\t convert numerator lattice ...")
numkeys = list(numlat.keys())
numkeys.sort()
conv_numlat = conv_lat_format(numkeys, numlat, filler, lm)
print("\t convert denominator lattice ...")
denkeys = list(den_wordseg.keys())
denkeys.sort()
denlat = conv_lat_format(denkeys, den_wordseg, filler, lm)
print("\t add numerator lattice into denominator lattice ...")
conv_denlat = add_numlat_into_denlat(conv_numlat, denlat)
numlatfile = "%s/%s.numlat" % (outdir, fileid)
denlatfile = "%s/%s.denlat" % (outdir, fileid)
print("\t write numerator lattice to %s ..." % numlatfile)
try:
output_lattice(conv_numlat, len(conv_numlat), numlatfile)
except IOError:
print("\t can't write numerator lattice to %s" % numlatfile)
continue
print("\t write denominator lattice to %s ...\n" % denlatfile)
try:
output_lattice(conv_denlat, len(conv_numlat), denlatfile)
except IOError:
print("\t can't write denominator lattice to %s" % denlatfile)
print("ALL DONE\n")
# If this script is being run from the command-line, do this
if __name__ == '__main__':
if len(sys.argv) != 9:
sys.stderr.write(
"Usage: %s LMFILE FILLERFILE FILELIST FILECOUNT FILEOFFSET DENLATDIR NUMLATDIR OUTLATDIR\n"
% (sys.argv[0]))
sys.exit(1)
command = ''
for argv in sys.argv:
command += argv + ' '
print("%s\n" % command)
lmfile, fillerfile, filelist, filecount, fileoffset, \
denlatdir, numlatdir, outdir = sys.argv[1:]
# load the filler dictionary
filler = load_lexicon(fillerfile)
# load the language model to score the word hypothesis in a lattice
lm = sphinxbase.NGramModel(lmfile)
# convert lattice and output to a file
write_lat(lm, filler, filelist, filecount, fileoffset, denlatdir,
numlatdir, outdir)
|
