gmm-decode-faster.cc File Reference

#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "gmm/am-diag-gmm.h"
#include "hmm/transition-model.h"
#include "fstext/fstext-lib.h"
#include "decoder/faster-decoder.h"
#include "gmm/decodable-am-diag-gmm.h"
#include "base/timer.h"
#include "lat/kaldi-lattice.h"

Include dependency graph for gmm-decode-faster.cc:

Go to the source code of this file.

Namespaces
	kaldi
	This code computes Goodness of Pronunciation (GOP) and extracts phone-level pronunciation feature for mispronunciations detection tasks, the reference:

Functions
fst::Fst< fst::StdArc > *	ReadNetwork (std::string filename)
int	main (int argc, char *argv[])

Function Documentation

main()

int main	(	int	argc,
		char *	argv[]
	)

Definition at line 67 of file gmm-decode-faster.cc.

References fst::AcousticLatticeScale(), fst::ConvertLattice(), FasterDecoder::Decode(), SequentialTableReader< Holder >::Done(), Timer::Elapsed(), SequentialTableReader< Holder >::FreeCurrent(), ParseOptions::GetArg(), FasterDecoder::GetBestPath(), fst::GetLinearSymbolSequence(), ParseOptions::GetOptArg(), rnnlm::i, TableWriter< Holder >::IsOpen(), KALDI_ERR, KALDI_LOG, KALDI_VLOG, KALDI_WARN, SequentialTableReader< Holder >::Key(), SequentialTableReader< Holder >::Next(), ParseOptions::NumArgs(), ParseOptions::PrintUsage(), FasterDecoder::ReachedFinal(), AmDiagGmm::Read(), ParseOptions::Read(), TransitionModel::Read(), kaldi::ReadNetwork(), FasterDecoderOptions::Register(), ParseOptions::Register(), fst::ScaleLattice(), SequentialTableReader< Holder >::Value(), LatticeWeightTpl< FloatType >::Value1(), LatticeWeightTpl< FloatType >::Value2(), words, and TableWriter< Holder >::Write().

                                 {
  try {
    using namespace kaldi;
    typedef kaldi::int32 int32;

    const char *usage =
        "Decode features using GMM-based model.\n"
        "Usage:  gmm-decode-faster [options] model-in fst-in features-rspecifier words-wspecifier [alignments-wspecifier [lattice-wspecifier]]\n"
        "Note: lattices, if output, will just be linear sequences; use gmm-latgen-faster\n"
        "  if you want \"real\" lattices.\n";
    ParseOptions po(usage);
    bool allow_partial = true;
    BaseFloat acoustic_scale = 0.1;
    
    std::string word_syms_filename;
    FasterDecoderOptions decoder_opts;
    decoder_opts.Register(&po, true);  // true == include obscure settings.
    po.Register("acoustic-scale", &acoustic_scale,
                "Scaling factor for acoustic likelihoods");
    po.Register("word-symbol-table", &word_syms_filename,
                "Symbol table for words [for debug output]");
    po.Register("allow-partial", &allow_partial,
                "Produce output even when final state was not reached");
    po.Read(argc, argv);

    if (po.NumArgs() < 4 || po.NumArgs() > 6) {
      po.PrintUsage();
      exit(1);
    }

    std::string model_rxfilename = po.GetArg(1),
        fst_rxfilename = po.GetArg(2),
        feature_rspecifier = po.GetArg(3),
        words_wspecifier = po.GetArg(4),
        alignment_wspecifier = po.GetOptArg(5),
        lattice_wspecifier = po.GetOptArg(6);

    TransitionModel trans_model;
    AmDiagGmm am_gmm;
    {
      bool binary;
      Input ki(model_rxfilename, &binary);
      trans_model.Read(ki.Stream(), binary);
      am_gmm.Read(ki.Stream(), binary);
    }

    Int32VectorWriter words_writer(words_wspecifier);

    Int32VectorWriter alignment_writer(alignment_wspecifier);

    CompactLatticeWriter clat_writer(lattice_wspecifier);

    fst::SymbolTable *word_syms = NULL;
    if (word_syms_filename != "") 
      if (!(word_syms = fst::SymbolTable::ReadText(word_syms_filename)))
        KALDI_ERR << "Could not read symbol table from file "
                   << word_syms_filename;

    SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);

    // It's important that we initialize decode_fst after feature_reader, as it
    // can prevent crashes on systems installed without enough virtual memory.
    // It has to do with what happens on UNIX systems if you call fork() on a
    // large process: the page-table entries are duplicated, which requires a
    // lot of virtual memory.
    fst::Fst<fst::StdArc> *decode_fst = ReadNetwork(fst_rxfilename);
    
    BaseFloat tot_like = 0.0;
    kaldi::int64 frame_count = 0;
    int num_success = 0, num_fail = 0;
    FasterDecoder decoder(*decode_fst, decoder_opts);

    Timer timer;

    for (; !feature_reader.Done(); feature_reader.Next()) {
      std::string key = feature_reader.Key();
      Matrix<BaseFloat> features (feature_reader.Value());
      feature_reader.FreeCurrent();
      if (features.NumRows() == 0) {
        KALDI_WARN << "Zero-length utterance: " << key;
        num_fail++;
        continue;
      }

      DecodableAmDiagGmmScaled gmm_decodable(am_gmm, trans_model, features,
                                             acoustic_scale);
      decoder.Decode(&gmm_decodable);

      fst::VectorFst<LatticeArc> decoded;  // linear FST.

      if ( (allow_partial || decoder.ReachedFinal())
           && decoder.GetBestPath(&decoded) ) {
        if (!decoder.ReachedFinal())
          KALDI_WARN << "Decoder did not reach end-state, "
                     << "outputting partial traceback since --allow-partial=true";
        num_success++;
        std::vector<int32> alignment;
        std::vector<int32> words;
        LatticeWeight weight;
        frame_count += features.NumRows();

        GetLinearSymbolSequence(decoded, &alignment, &words, &weight);

        words_writer.Write(key, words);
        if (alignment_writer.IsOpen())
          alignment_writer.Write(key, alignment);
          
        if (lattice_wspecifier != "") {
          // We'll write the lattice without acoustic scaling.
          if (acoustic_scale != 0.0)
            fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale),
                              &decoded);
          fst::VectorFst<CompactLatticeArc> clat;
          ConvertLattice(decoded, &clat, true);
          clat_writer.Write(key, clat);
        }
          
        if (word_syms != NULL) {
          std::cerr << key << ' ';
          for (size_t i = 0; i < words.size(); i++) {
            std::string s = word_syms->Find(words[i]);
            if (s == "")
              KALDI_ERR << "Word-id " << words[i] <<" not in symbol table.";
            std::cerr << s << ' ';
          }
          std::cerr << '\n';
        }
        BaseFloat like = -weight.Value1() -weight.Value2();
        tot_like += like;
        KALDI_LOG << "Log-like per frame for utterance " << key << " is "
                  << (like / features.NumRows()) << " over "
                  << features.NumRows() << " frames.";
        KALDI_VLOG(2) << "Cost for utterance " << key << " is "
                      << weight.Value1() << " + " << weight.Value2();
      } else {
        num_fail++;
        KALDI_WARN << "Did not successfully decode utterance " << key
                   << ", len = " << features.NumRows();
      }
    }

    double elapsed = timer.Elapsed();
    KALDI_LOG << "Time taken [excluding initialization] "<< elapsed
              << "s: real-time factor assuming 100 frames/sec is "
              << (elapsed*100.0/frame_count);
    KALDI_LOG << "Done " << num_success << " utterances, failed for "
              << num_fail;
    KALDI_LOG << "Overall log-likelihood per frame is " << (tot_like/frame_count) << " over "
              << frame_count<<" frames.";

    delete word_syms;    
    delete decode_fst;
    return (num_success != 0 ? 0 : 1);
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}