All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Modules Pages
gmm-decode-faster-regtree-mllr.cc File Reference
#include <string>
#include <vector>
#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "gmm/am-diag-gmm.h"
#include "hmm/transition-model.h"
#include "transform/regression-tree.h"
#include "transform/regtree-mllr-diag-gmm.h"
#include "fstext/fstext-lib.h"
#include "decoder/faster-decoder.h"
#include "transform/decodable-am-diag-gmm-regtree.h"
#include "base/timer.h"
#include "lat/kaldi-lattice.h"
Include dependency graph for gmm-decode-faster-regtree-mllr.cc:

Go to the source code of this file.

Classes

struct  DecodeInfo
 

Functions

bool DecodeUtterance (kaldi::FasterDecoder *decoder, kaldi::DecodableInterface *decodable, DecodeInfo *info, const string &uttid, int32 num_frames, BaseFloat *total_like)
 
int main (int argc, char *argv[])
 

Function Documentation

bool DecodeUtterance ( kaldi::FasterDecoder decoder,
kaldi::DecodableInterface decodable,
DecodeInfo info,
const string &  uttid,
int32  num_frames,
BaseFloat *  total_like 
)

Definition at line 69 of file gmm-decode-faster-regtree-mllr.cc.

References DecodeInfo::alignment_writer, DecodeInfo::allow_partial, FasterDecoder::Decode(), FasterDecoder::GetBestPath(), fst::GetLinearSymbolSequence(), rnnlm::i, TableWriter< Holder >::IsOpen(), KALDI_ERR, KALDI_LOG, KALDI_WARN, FasterDecoder::ReachedFinal(), DecodeInfo::word_syms, words, DecodeInfo::words_writer, and TableWriter< Holder >::Write().

74  {
75  decoder->Decode(decodable);
76  KALDI_LOG << "Length of file is " << num_frames;;
77 
78  VectorFst<LatticeArc> decoded; // linear FST.
79  if ( (info->allow_partial || decoder->ReachedFinal())
80  && decoder->GetBestPath(&decoded) ) {
81  if (!decoder->ReachedFinal())
82  KALDI_WARN << "Decoder did not reach end-state, outputting partial "
83  "traceback.";
84 
85  vector<kaldi::int32> alignment, words;
86  LatticeWeight weight;
87  GetLinearSymbolSequence(decoded, &alignment, &words, &weight);
88 
89  info->words_writer.Write(uttid, words);
90  if (info->alignment_writer.IsOpen())
91  info->alignment_writer.Write(uttid, alignment);
92  if (info->word_syms != NULL) {
93  std::ostringstream ss;
94  ss << uttid << ' ';
95  for (size_t i = 0; i < words.size(); i++) {
96  string s = info->word_syms->Find(words[i]);
97  if (s == "")
98  KALDI_ERR << "Word-id " << words[i] << " not in symbol table.";
99  ss << s << ' ';
100  }
101  ss << '\n';
102  KALDI_LOG << ss.str();
103  }
104 
105  BaseFloat like = -weight.Value1() -weight.Value2();
106  KALDI_LOG << "Log-like per frame = " << (like/num_frames);
107  (*total_like) += like;
108  return true;
109  } else {
110  KALDI_WARN << "Did not successfully decode utterance, length = "
111  << num_frames;
112  return false;
113  }
114 }
int32 words[kMaxOrder]
void Write(const std::string &key, const T &value) const
const kaldi::Int32VectorWriter & alignment_writer
bool IsOpen() const
void Decode(DecodableInterface *decodable)
bool GetBestPath(fst::MutableFst< LatticeArc > *fst_out, bool use_final_probs=true)
GetBestPath gets the decoding traceback.
float BaseFloat
Definition: kaldi-types.h:29
#define KALDI_ERR
Definition: kaldi-error.h:127
bool ReachedFinal()
Returns true if a final state was active on the last frame.
#define KALDI_WARN
Definition: kaldi-error.h:130
bool GetLinearSymbolSequence(const Fst< Arc > &fst, vector< I > *isymbols_out, vector< I > *osymbols_out, typename Arc::Weight *tot_weight_out)
GetLinearSymbolSequence gets the symbol sequence from a linear FST.
fst::SymbolTable * word_syms
LatticeWeightTpl< BaseFloat > LatticeWeight
#define KALDI_LOG
Definition: kaldi-error.h:133
const kaldi::Int32VectorWriter & words_writer
int main ( int  argc,
char *  argv[] 
)

Definition at line 116 of file gmm-decode-faster-regtree-mllr.cc.

References kaldi::DecodeUtterance(), SequentialTableReader< Holder >::Done(), Timer::Elapsed(), SequentialTableReader< Holder >::FreeCurrent(), ParseOptions::GetArg(), ParseOptions::GetOptArg(), RandomAccessTableReaderMapped< Holder >::HasKey(), KALDI_ERR, KALDI_LOG, KALDI_WARN, SequentialTableReader< Holder >::Key(), SequentialTableReader< Holder >::Next(), ParseOptions::NumArgs(), DecodableAmDiagGmmUnmapped::NumFramesReady(), DecodableAmDiagGmmRegtreeMllr::NumFramesReady(), ParseOptions::PrintUsage(), RegressionTree::Read(), AmDiagGmm::Read(), ParseOptions::Read(), TransitionModel::Read(), fst::ReadFstKaldi(), FasterDecoderOptions::Register(), ParseOptions::Register(), Input::Stream(), SequentialTableReader< Holder >::Value(), and RandomAccessTableReaderMapped< Holder >::Value().

116  {
117  try {
118  using namespace kaldi;
119  typedef kaldi::int32 int32;
120 
121  const char *usage = "Decode features using GMM-based model.\n"
122  "Usage: gmm-decode-faster-regtree-mllr [options] model-in fst-in "
123  "regtree-in features-rspecifier transforms-rspecifier "
124  "words-wspecifier [alignments-wspecifier]\n";
125  ParseOptions po(usage);
126  bool binary = true;
127  bool allow_partial = true;
128  BaseFloat acoustic_scale = 0.1;
129 
130  std::string word_syms_filename, utt2spk_rspecifier;
131  FasterDecoderOptions decoder_opts;
132  decoder_opts.Register(&po, true); // true == include obscure settings.
133  po.Register("utt2spk", &utt2spk_rspecifier, "rspecifier for utterance to "
134  "speaker map");
135  po.Register("binary", &binary, "Write output in binary mode");
136  po.Register("acoustic-scale", &acoustic_scale,
137  "Scaling factor for acoustic likelihoods");
138  po.Register("word-symbol-table", &word_syms_filename,
139  "Symbol table for words [for debug output]");
140  po.Register("allow-partial", &allow_partial,
141  "Produce output even when final state was not reached");
142  po.Read(argc, argv);
143 
144  if (po.NumArgs() < 6 || po.NumArgs() > 7) {
145  po.PrintUsage();
146  exit(1);
147  }
148 
149  std::string model_in_filename = po.GetArg(1),
150  fst_in_filename = po.GetArg(2),
151  regtree_filename = po.GetArg(3),
152  feature_rspecifier = po.GetArg(4),
153  xforms_rspecifier = po.GetArg(5),
154  words_wspecifier = po.GetArg(6),
155  alignment_wspecifier = po.GetOptArg(7);
156 
157  TransitionModel trans_model;
158  AmDiagGmm am_gmm;
159  {
160  bool binary_read;
161  Input ki(model_in_filename, &binary_read);
162  trans_model.Read(ki.Stream(), binary_read);
163  am_gmm.Read(ki.Stream(), binary_read);
164  }
165 
166  VectorFst<StdArc> *decode_fst = fst::ReadFstKaldi(fst_in_filename);
167 
168  RegressionTree regtree;
169  {
170  bool binary_read;
171  Input in(regtree_filename, &binary_read);
172  regtree.Read(in.Stream(), binary_read, am_gmm);
173  }
174 
175  RandomAccessRegtreeMllrDiagGmmReaderMapped mllr_reader(xforms_rspecifier,
176  utt2spk_rspecifier);
177 
178  Int32VectorWriter words_writer(words_wspecifier);
179 
180  Int32VectorWriter alignment_writer(alignment_wspecifier);
181 
182  fst::SymbolTable *word_syms = NULL;
183  if (word_syms_filename != "") {
184  word_syms = fst::SymbolTable::ReadText(word_syms_filename);
185  if (!word_syms) {
186  KALDI_ERR << "Could not read symbol table from file "
187  << word_syms_filename;
188  }
189  }
190 
191  BaseFloat tot_like = 0.0;
192  kaldi::int64 frame_count = 0;
193  int num_success = 0, num_fail = 0;
194  FasterDecoder decoder(*decode_fst, decoder_opts);
195 
196  Timer timer;
197 
198  DecodeInfo decode_info(am_gmm, trans_model, &decoder, acoustic_scale,
199  allow_partial, words_writer, alignment_writer,
200  word_syms);
201 
202  SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
203  for (; !feature_reader.Done(); feature_reader.Next()) {
204  string utt = feature_reader.Key();
205 
206  Matrix<BaseFloat> features(feature_reader.Value());
207  feature_reader.FreeCurrent();
208  if (features.NumRows() == 0) {
209  KALDI_WARN << "Zero-length utterance: " << utt;
210  num_fail++;
211  continue;
212  }
213 
214  if (!mllr_reader.HasKey(utt)) { // Decode without MLLR if none found
215  KALDI_WARN << "No MLLR transform for key " << utt <<
216  ", decoding without MLLR.";
217  kaldi::DecodableAmDiagGmmScaled gmm_decodable(am_gmm, trans_model,
218  features,
219  acoustic_scale);
220  if (DecodeUtterance(&decoder, &gmm_decodable, &decode_info,
221  utt, features.NumRows(), &tot_like)) {
222  frame_count += gmm_decodable.NumFramesReady();
223  num_success++;
224  } else {
225  num_fail++;
226  }
227  continue;
228  }
229 
230  // If found, load the transforms for the current utterance.
231  const RegtreeMllrDiagGmm &mllr = mllr_reader.Value(utt);
232  kaldi::DecodableAmDiagGmmRegtreeMllr gmm_decodable(am_gmm, trans_model,
233  features, mllr,
234  regtree,
235  acoustic_scale);
236  if (DecodeUtterance(&decoder, &gmm_decodable, &decode_info,
237  utt, features.NumRows(), &tot_like)) {
238  frame_count += gmm_decodable.NumFramesReady();
239  num_success++;
240  } else {
241  num_fail++;
242  }
243  } // end looping over all utterances
244 
245  double elapsed = timer.Elapsed();
246  KALDI_LOG << "Time taken [excluding initialization] " << elapsed
247  << "s: real-time factor assuming 100 frames/sec is "
248  << (elapsed * 100.0 / frame_count);
249  KALDI_LOG << "Done " << num_success << " utterances, failed for "
250  << num_fail;
251  KALDI_LOG << "Overall log-likelihood per frame is "
252  << (tot_like / frame_count) << " over " << frame_count
253  << " frames.";
254 
255  delete decode_fst;
256  if (num_success != 0)
257  return 0;
258  else
259  return 1;
260  }
261  catch(const std::exception &e) {
262  std::cerr << e.what();
263  return -1;
264  }
265 }
bool DecodeUtterance(LatticeBiglmFasterDecoder &decoder, DecodableInterface &decodable, const TransitionModel &trans_model, const fst::SymbolTable *word_syms, std::string utt, double acoustic_scale, bool determinize, bool allow_partial, Int32VectorWriter *alignment_writer, Int32VectorWriter *words_writer, CompactLatticeWriter *compact_lattice_writer, LatticeWriter *lattice_writer, double *like_ptr)
void Read(std::istream &in, bool binary, const AmDiagGmm &am)
Relabels neural network egs with the read pdf-id alignments.
Definition: chain.dox:20
An MLLR mean transformation is an affine transformation of Gaussian means.
This class is for when you are reading something in random access, but it may actually be stored per-...
Definition: kaldi-table.h:431
A templated class for writing objects to an archive or script file; see The Table concept...
Definition: kaldi-table.h:366
float BaseFloat
Definition: kaldi-types.h:29
The class ParseOptions is for parsing command-line options; see Parsing command-line options for more...
Definition: parse-options.h:36
A regression tree is a clustering of Gaussian densities in an acoustic model, such that the group of ...
void Read(std::istream &is, bool binary)
A templated class for reading objects sequentially from an archive or script file; see The Table conc...
Definition: kaldi-table.h:287
#define KALDI_ERR
Definition: kaldi-error.h:127
#define KALDI_WARN
Definition: kaldi-error.h:130
void Register(OptionsItf *opts, bool full)
void ReadFstKaldi(std::istream &is, bool binary, VectorFst< Arc > *fst)
double Elapsed() const
Returns time in seconds.
Definition: timer.h:74
#define KALDI_LOG
Definition: kaldi-error.h:133
void Read(std::istream &in_stream, bool binary)
Definition: am-diag-gmm.cc:147