gmm-est-regtree-fmllr-ali.cc
7.82 KB
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
// gmmbin/gmm-est-regtree-fmllr-ali.cc
// Copyright 2009-2011 Saarland University; Microsoft Corporation
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include <string>
using std::string;
#include <vector>
using std::vector;
#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "gmm/am-diag-gmm.h"
#include "hmm/transition-model.h"
#include "transform/regtree-fmllr-diag-gmm.h"
int main(int argc, char *argv[]) {
try {
typedef kaldi::int32 int32;
using namespace kaldi;
const char *usage =
"Compute FMLLR transforms per-utterance (default) or per-speaker for "
"the supplied set of speakers (spk2utt option). Note: writes RegtreeFmllrDiagGmm objects\n"
"Usage: gmm-est-regtree-fmllr-ali [options] <model-in> <feature-rspecifier> "
"<alignments-rspecifier> <regression-tree> <transforms-wspecifier>\n";
ParseOptions po(usage);
string spk2utt_rspecifier;
bool binary = true;
po.Register("spk2utt", &spk2utt_rspecifier, "rspecifier for speaker to "
"utterance-list map");
po.Register("binary", &binary, "Write output in binary mode");
// register other modules
RegtreeFmllrOptions opts;
opts.Register(&po);
po.Read(argc, argv);
if (po.NumArgs() != 5) {
po.PrintUsage();
exit(1);
}
string model_filename = po.GetArg(1),
feature_rspecifier = po.GetArg(2),
alignments_rspecifier = po.GetArg(3),
regtree_filename = po.GetArg(4),
xforms_wspecifier = po.GetArg(5);
RandomAccessInt32VectorReader alignments_reader(alignments_rspecifier);
RegtreeFmllrDiagGmmWriter fmllr_writer(xforms_wspecifier);
AmDiagGmm am_gmm;
TransitionModel trans_model;
{
bool binary;
Input ki(model_filename, &binary);
trans_model.Read(ki.Stream(), binary);
am_gmm.Read(ki.Stream(), binary);
}
RegressionTree regtree;
{
bool binary;
Input in(regtree_filename, &binary);
regtree.Read(in.Stream(), binary, am_gmm);
}
RegtreeFmllrDiagGmm fmllr_xforms;
RegtreeFmllrDiagGmmAccs fmllr_accs;
fmllr_accs.Init(regtree.NumBaseclasses(), am_gmm.Dim());
double tot_like = 0.0;
kaldi::int64 tot_t = 0;
int32 num_done = 0, num_no_alignment = 0, num_other_error = 0;
double tot_objf_impr = 0.0, tot_t_objf = 0.0;
if (spk2utt_rspecifier != "") { // per-speaker adaptation
SequentialTokenVectorReader spk2utt_reader(spk2utt_rspecifier);
RandomAccessBaseFloatMatrixReader feature_reader(feature_rspecifier);
for (; !spk2utt_reader.Done(); spk2utt_reader.Next()) {
string spk = spk2utt_reader.Key();
fmllr_accs.SetZero();
const vector<string> &uttlist = spk2utt_reader.Value();
for (vector<string>::const_iterator utt_itr = uttlist.begin(),
itr_end = uttlist.end(); utt_itr != itr_end; ++utt_itr) {
if (!feature_reader.HasKey(*utt_itr)) {
KALDI_WARN << "Did not find features for utterance " << *utt_itr;
continue;
}
if (!alignments_reader.HasKey(*utt_itr)) {
KALDI_WARN << "Did not find aligned transcription for utterance "
<< *utt_itr;
num_no_alignment++;
continue;
}
const Matrix<BaseFloat> &feats = feature_reader.Value(*utt_itr);
const vector<int32> &alignment = alignments_reader.Value(*utt_itr);
if (static_cast<int32>(alignment.size()) != feats.NumRows()) {
KALDI_WARN << "Alignments has wrong size " << (alignment.size())
<< " vs. " << (feats.NumRows());
num_other_error++;
continue;
}
BaseFloat file_like = 0.0;
for (size_t i = 0; i < alignment.size(); i++) {
int32 pdf_id = trans_model.TransitionIdToPdf(alignment[i]);
file_like += fmllr_accs.AccumulateForGmm(regtree, am_gmm,
feats.Row(i), pdf_id, 1.0);
}
KALDI_VLOG(2) << "Average like for this file is " << (file_like
/ alignment.size()) << " over " << alignment.size()
<< " frames.\n";
tot_like += file_like;
tot_t += alignment.size();
num_done++;
if (num_done % 10 == 0) KALDI_VLOG(1)
<< "Avg like per frame so far is " << (tot_like / tot_t) << '\n';
} // end looping over all utterances of the current speaker
BaseFloat objf_impr, t;
fmllr_accs.Update(regtree, opts, &fmllr_xforms, &objf_impr, &t);
KALDI_LOG << "fMLLR objf improvement for speaker " << spk << " is "
<< (objf_impr/(t+1.0e-10)) << " per frame over " << t
<< " frames.";
tot_objf_impr += objf_impr;
tot_t_objf += t;
fmllr_writer.Write(spk, fmllr_xforms);
} // end looping over speakers
} else { // per-utterance adaptation
SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
for (; !feature_reader.Done(); feature_reader.Next()) {
string key = feature_reader.Key();
if (!alignments_reader.HasKey(key)) {
KALDI_WARN << "Did not find aligned transcription for utterance "
<< key;
num_no_alignment++;
continue;
}
const Matrix<BaseFloat> &feats = feature_reader.Value();
const vector<int32> &alignment = alignments_reader.Value(key);
if (static_cast<int32>(alignment.size()) != feats.NumRows()) {
KALDI_WARN << "Alignments has wrong size " << (alignment.size())
<< " vs. " << (feats.NumRows());
num_other_error++;
continue;
}
num_done++;
BaseFloat file_like = 0.0;
fmllr_accs.SetZero();
for (size_t i = 0; i < alignment.size(); i++) {
int32 pdf_id = trans_model.TransitionIdToPdf(alignment[i]);
file_like += fmllr_accs.AccumulateForGmm(regtree, am_gmm,
feats.Row(i), pdf_id, 1.0);
}
KALDI_VLOG(2) << "Average like for this file is " << (file_like
/ alignment.size()) << " over " << alignment.size() << " frames.";
tot_like += file_like;
tot_t += alignment.size();
if (num_done % 10 == 0) KALDI_VLOG(1)
<< "Avg like per frame so far is " << (tot_like / tot_t);
BaseFloat objf_impr, t;
fmllr_accs.Update(regtree, opts, &fmllr_xforms, &objf_impr, &t);
KALDI_LOG << "fMLLR objf improvement for utterance " << key << " is "
<< (objf_impr/(t+1.0e-10)) << " per frame over " << t
<< " frames.";
tot_objf_impr += objf_impr;
tot_t_objf += t;
fmllr_writer.Write(feature_reader.Key(), fmllr_xforms);
}
}
KALDI_LOG << "Overall objf improvement from fMLLR is "
<< (tot_objf_impr/tot_t_objf)
<< " per frame over " << tot_t_objf << " frames.";
KALDI_LOG << "Done " << num_done << " files, " << num_no_alignment
<< " with no alignments, " << num_other_error
<< " with other errors.";
KALDI_LOG << "Overall acoustic like per frame = " << (tot_like / tot_t)
<< " over " << tot_t << " frames.";
return 0;
} catch(const std::exception &e) {
std::cerr << e.what();
return -1;
}
}