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
|
const char *help = "\
speech_hmm_init (c) Samy Bengio & Co 2001\n\
\n\
This program is used to do speech recognition \n";
#include <torch/LexiconInfo.h>
#include <torch/EMTrainer.h>
#include <torch/HMM.h>
#include <torch/KMeans.h>
#include <torch/MatDataSet.h>
#include <torch/CmdLine.h>
#include <torch/NLLMeasurer.h>
#include <torch/Random.h>
#include <torch/DiskHTKDataSet.h>
#include <torch/HTKDataSet.h>
#include <torch/SpeechHMM.h>
#include <torch/string_utils.h>
#include <torch/FileListCmdOption.h>
using namespace Torch;
#ifdef _MSC_VER
#include "speech_include.cpp"
#else
#include "speech_include.cc"
#endif
int main(int argc, char **argv)
{
char* phoneme_name;
char* lex_name;
int n_gaussians;
int n_states;
char* file_states;
char* silence_name;
char* silence_phone;
char* sp_name;
char* sp_phone;
real accuracy;
real threshold;
int max_iter;
real prior;
int max_load;
int the_seed;
char *dir_name;
char *model_file;
bool htk_model;
bool train_separate;
bool disk;
bool is_word;
bool add_sil_to_targets;
bool full_connect;
Allocator *allocator = new Allocator;
FileListCmdOption input_file_list("file name", "the list of inputs files or one data file");
input_file_list.isArgument(true);
FileListCmdOption target_file_list("file name", "the list of target files or one target file");
target_file_list.isArgument(true);
//===============================================================
//=================== The command-line ==========================
//===============================================================
// Construct the command line
CmdLine cmd;
// Put the help line at the beginning
cmd.info(help);
// Train mode
cmd.addText("\nArguments:");
cmd.addSCmdArg("phoneme_name", &phoneme_name, "the list of phonemes file");
cmd.addSCmdArg("lex_name", &lex_name, "the lexicon file");
cmd.addCmdOption(&input_file_list);
cmd.addCmdOption(&target_file_list);
cmd.addText("\nModel Options:");
cmd.addBCmdOption("-word", &is_word, false, "targets are in word format");
cmd.addICmdOption("-n_gaussians", &n_gaussians, 10, "number of Gaussians");
cmd.addICmdOption("-n_states", &n_states, 5, "number of states");
cmd.addSCmdOption("-file_states", &file_states, "", "file containing n_states per model");
cmd.addSCmdOption("-silence_word", &silence_name,"sil", "name of silence word");
cmd.addSCmdOption("-silence_phone", &silence_phone,"h#", "name of silence phone");
cmd.addSCmdOption("-sp_word", &sp_name,"", "name of short pause word");
cmd.addSCmdOption("-sp_phone", &sp_phone,"", "name of short pause phoneme");
cmd.addBCmdOption("-add_sil_to_targets", &add_sil_to_targets, false, "add silence at begining of targets");
cmd.addBCmdOption("-full_connect", &full_connect, false, "models are full connect");
cmd.addText("\nLearning Options:");
cmd.addBCmdOption("-train_separate", &train_separate, false, "first train separate models");
cmd.addRCmdOption("-threshold", &threshold, 0.001, "variance threshold");
cmd.addRCmdOption("-prior", &prior, 0.001, "prior on the weights");
cmd.addICmdOption("-iter", &max_iter, 25, "max number of init iterations");
cmd.addRCmdOption("-e", &accuracy, 0.00001, "end accuracy");
cmd.addText("\nMisc Options:");
cmd.addICmdOption("-seed", &the_seed, -1, "the random seed");
cmd.addICmdOption("-load", &max_load, -1, "max number of examples to load for train");
cmd.addSCmdOption("-dir", &dir_name, ".", "directory to save measures");
cmd.addSCmdOption("-save", &model_file, "", "the model file");
cmd.addBCmdOption("-htk_model", &htk_model, false, "save in HTK model format");
cmd.addBCmdOption("-disk", &disk, false, "keep data on disk");
// Read the command line
cmd.read(argc, argv);
cmd.setWorkingDirectory(dir_name);
DiskXFile::setBigEndianMode();
//====================================================================
//=================== Training Mode =================================
//====================================================================
if(the_seed == -1)
Random::seed();
else
Random::manualSeed((long)the_seed);
// read lexicon
if (strlen(sp_phone) == 0)
sp_phone = NULL;
if (strlen(silence_phone) == 0)
silence_phone = NULL;
LexiconInfo lex(phoneme_name,silence_phone,sp_phone,lex_name,add_sil_to_targets ? silence_name : NULL,NULL);
//====================================================================
//=================== Create the DataSet ... =========================
//====================================================================
// some basic tests on the files
if (input_file_list.n_files != target_file_list.n_files) {
error("the input and target files should have the same number of files (%d != %d)\n",input_file_list.n_files,target_file_list.n_files);
}
for (int i=0;i<input_file_list.n_files;i++) {
char* si = strRemoveSuffix(input_file_list.file_names[i]);
char* st = strRemoveSuffix(target_file_list.file_names[i]);
if (strcmp(strBaseName(si),strBaseName(st)))
warning("input file (%s) do not correspond to target file (%s)",input_file_list.file_names[i],target_file_list.file_names[i]);
free(si);
free(st);
}
DataSet* data;
if (disk)
data = (DataSet*)new(allocator) DiskHTKDataSet(input_file_list.file_names,target_file_list.file_names, input_file_list.n_files, true, max_load, &lex,is_word);
else
data = (DataSet*)new(allocator) HTKDataSet(input_file_list.file_names,target_file_list.file_names, input_file_list.n_files, true, max_load, &lex,is_word);
//=================== Create the HMM... =========================
real* thresh = (real*)allocator->alloc(data->n_inputs*sizeof(real));
initializeThreshold(data,thresh,threshold);
int n_phonemes = lex.phone_info->n_phones;
int* states = read_states(file_states,n_phonemes,n_states,
lex.phone_info->pause_index,allocator);
// create the GMM
DiagonalGMM*** gmms = (DiagonalGMM ***)allocator->alloc(sizeof(DiagonalGMM**)*n_phonemes);
EMTrainer** hmm_trainer = (EMTrainer **)allocator->alloc(sizeof(EMTrainer*)*n_phonemes);
HMM** hmm = (HMM**)allocator->alloc(sizeof(HMM*)*n_phonemes);
KMeans*** kmeans = (KMeans***)allocator->alloc(sizeof(KMeans**)*n_phonemes);
EMTrainer*** kmeans_trainer = (EMTrainer***)allocator->alloc(sizeof(EMTrainer**)*n_phonemes);
real*** transitions = (real***)allocator->alloc(n_phonemes*sizeof(real**));
for (int i=0;i<n_phonemes;i++) {
gmms[i] = (DiagonalGMM**)allocator->alloc(sizeof(DiagonalGMM*)*states[i]);
kmeans[i] = (KMeans**)allocator->alloc(sizeof(KMeans*)*states[i]);
kmeans_trainer[i] = (EMTrainer**)allocator->alloc(sizeof(EMTrainer*)*states[i]);
for (int j=1;j<states[i]-1;j++) {
// create a KMeans object to initialize the GMM
kmeans[i][j] = new(allocator)KMeans(data->n_inputs, n_gaussians);
kmeans[i][j]->setVarThreshold(thresh);
// the kmeans trainer
kmeans_trainer[i][j] = new(allocator) EMTrainer(kmeans[i][j]);
kmeans_trainer[i][j]->setROption("end accuracy", accuracy);
kmeans_trainer[i][j]->setIOption("max iter", max_iter);
gmms[i][j] = new(allocator)DiagonalGMM(data->n_inputs,n_gaussians,kmeans_trainer[i][j]);
// set the training options
gmms[i][j]->setVarThreshold(thresh);
gmms[i][j]->setROption("prior weights",prior);
}
gmms[i][0] = NULL;
gmms[i][states[i]-1] = NULL;
// the transition table probability: left-right topology or full connect
transitions[i] = (real**)allocator->alloc(states[i]*sizeof(real*));
for (int j=0;j<states[i];j++) {
transitions[i][j] = (real*)allocator->alloc(states[i]*sizeof(real));
}
setTransitions(transitions[i], states[i], i, lex.phone_info->pause_index, lex.phone_info->sil_index, full_connect);
hmm[i] = new (allocator)HMM(states[i],(Distribution**)gmms[i],transitions[i]);
hmm[i]->setROption("prior transitions",prior);
hmm[i]->setBOption("linear segmentation", true);
hmm_trainer[i] = new(allocator)EMTrainer(hmm[i]);
hmm_trainer[i]->setIOption("max iter", max_iter);
hmm_trainer[i]->setROption("end accuracy", accuracy);
}
EMTrainer** hmm_trainer_ = NULL;
if (train_separate)
hmm_trainer_ = hmm_trainer;
SpeechHMM shmm(n_phonemes,hmm,&lex,hmm_trainer_);
shmm.setBOption("targets are phonemes",!is_word);
shmm.setDataSet(data);
if(strcmp(model_file, "")) {
if (htk_model) {
save_htk_model(cmd.getPath(model_file),&shmm,lex.phone_info->phone_names);
} else {
DiskXFile model_(cmd.getPath(model_file), "w");
cmd.saveXFile(&model_);
model_.taggedWrite(&n_gaussians, sizeof(int), 1, "n_gaussians");
model_.taggedWrite(&n_phonemes, sizeof(int), 1, "n_phonemes");
model_.taggedWrite(states, sizeof(int), n_phonemes, "n_states");
model_.taggedWrite(&data->n_inputs, sizeof(int), 1, "n_inputs");
shmm.saveXFile(&model_);
}
}
delete allocator;
return(0);
}
|
