gmm-latgen-biglm-faster.cc File Reference

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

Include dependency graph for gmm-latgen-biglm-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
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)
int	main (int argc, char *argv[])

Function Documentation

main()

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

Definition at line 143 of file gmm-latgen-biglm-faster.cc.

References fst::ApplyProbabilityScale(), fst::CastOrConvertToVectorFst(), kaldi::ClassifyRspecifier(), kaldi::DecodeUtterance(), LatticeFasterDecoderConfig::determinize_lattice, SequentialTableReader< Holder >::Done(), Timer::Elapsed(), SequentialTableReader< Holder >::FreeCurrent(), ParseOptions::GetArg(), ParseOptions::GetOptArg(), RandomAccessTableReader< Holder >::HasKey(), KALDI_ERR, KALDI_LOG, KALDI_WARN, SequentialTableReader< Holder >::Key(), kaldi::kNoRspecifier, SequentialTableReader< Holder >::Next(), ParseOptions::NumArgs(), MatrixBase< Real >::NumRows(), TableWriter< Holder >::Open(), ParseOptions::PrintUsage(), AmDiagGmm::Read(), ParseOptions::Read(), TransitionModel::Read(), fst::ReadFstKaldi(), fst::ReadFstKaldiGeneric(), LatticeFasterDecoderConfig::Register(), ParseOptions::Register(), Input::Stream(), RandomAccessTableReader< Holder >::Value(), and SequentialTableReader< Holder >::Value().

                                 {
  try {
    using namespace kaldi;
    typedef kaldi::int32 int32;
    using fst::SymbolTable;
    using fst::VectorFst;
    using fst::Fst;
    using fst::StdArc;
    using fst::ReadFstKaldi;

    const char *usage =
        "Generate lattices using GMM-based model.\n"
        "User supplies LM used to generate decoding graph, and desired LM;\n"
        "this decoder applies the difference during decoding\n"
        "Usage: gmm-latgen-biglm-faster [options] model-in (fst-in|fsts-rspecifier) "
        "oldlm-fst-in newlm-fst-in features-rspecifier"
        " lattice-wspecifier [ words-wspecifier [alignments-wspecifier] ]\n";
    ParseOptions po(usage);
    Timer timer;
    bool allow_partial = false;
    BaseFloat acoustic_scale = 0.1;
    LatticeBiglmFasterDecoderConfig config;
    
    std::string word_syms_filename;
    config.Register(&po);
    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, "If true, produce output even if end state was not reached.");
    
    po.Read(argc, argv);

    if (po.NumArgs() < 6 || po.NumArgs() > 8) {
      po.PrintUsage();
      exit(1);
    }
    
    std::string model_in_filename = po.GetArg(1),
        fst_in_str = po.GetArg(2),
        old_lm_fst_rxfilename = po.GetArg(3),
        new_lm_fst_rxfilename = po.GetArg(4),
        feature_rspecifier = po.GetArg(5),
        lattice_wspecifier = po.GetArg(6),
        words_wspecifier = po.GetOptArg(7),
        alignment_wspecifier = po.GetOptArg(8);
    
    TransitionModel trans_model;
    AmDiagGmm am_gmm;
    {
      bool binary;
      Input ki(model_in_filename, &binary);
      trans_model.Read(ki.Stream(), binary);
      am_gmm.Read(ki.Stream(), binary);
    }
    
    VectorFst<StdArc> *old_lm_fst = fst::CastOrConvertToVectorFst(
        fst::ReadFstKaldiGeneric(old_lm_fst_rxfilename));
    ApplyProbabilityScale(-1.0, old_lm_fst); // Negate old LM probs...
    
    VectorFst<StdArc> *new_lm_fst = fst::CastOrConvertToVectorFst(
        fst::ReadFstKaldiGeneric(new_lm_fst_rxfilename));

    fst::BackoffDeterministicOnDemandFst<StdArc> old_lm_dfst(*old_lm_fst);
    fst::BackoffDeterministicOnDemandFst<StdArc> new_lm_dfst(*new_lm_fst);
    fst::ComposeDeterministicOnDemandFst<StdArc> compose_dfst(&old_lm_dfst,
                                                              &new_lm_dfst);
    fst::CacheDeterministicOnDemandFst<StdArc> cache_dfst(&compose_dfst);

    bool determinize = config.determinize_lattice;
    CompactLatticeWriter compact_lattice_writer;
    LatticeWriter lattice_writer;
    if (! (determinize ? compact_lattice_writer.Open(lattice_wspecifier)
           : lattice_writer.Open(lattice_wspecifier)))
      KALDI_ERR << "Could not open table for writing lattices: "
                 << lattice_wspecifier;

    Int32VectorWriter words_writer(words_wspecifier);

    Int32VectorWriter alignment_writer(alignment_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;

    double tot_like = 0.0;
    kaldi::int64 frame_count = 0;
    int num_success = 0, num_fail = 0;


    if (ClassifyRspecifier(fst_in_str, NULL, NULL) == kNoRspecifier) {
      SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
      // Input FST is just one FST, not a table of FSTs.
      Fst<StdArc> *decode_fst = fst::ReadFstKaldiGeneric(fst_in_str);

      {
        LatticeBiglmFasterDecoder decoder(*decode_fst, config, &cache_dfst);
    
        for (; !feature_reader.Done(); feature_reader.Next()) {
          std::string utt = feature_reader.Key();
          Matrix<BaseFloat> features (feature_reader.Value());
          feature_reader.FreeCurrent();
          if (features.NumRows() == 0) {
            KALDI_WARN << "Zero-length utterance: " << utt;
            num_fail++;
            continue;
          }
      
          DecodableAmDiagGmmScaled gmm_decodable(am_gmm, trans_model, features,
                                                 acoustic_scale);


          double like;
          if (DecodeUtterance(decoder, gmm_decodable, trans_model, word_syms,
                              utt, acoustic_scale, determinize, allow_partial,
                              &alignment_writer, &words_writer,
                              &compact_lattice_writer, &lattice_writer,
                              &like)) {
            tot_like += like;
            frame_count += features.NumRows();
            num_success++;
          } else num_fail++;
        }
      }
      delete decode_fst; // delete this only after decoder goes out of scope.
    } else { // We have different FSTs for different utterances.
      SequentialTableReader<fst::VectorFstHolder> fst_reader(fst_in_str);
      RandomAccessBaseFloatMatrixReader feature_reader(feature_rspecifier);          
      for (; !fst_reader.Done(); fst_reader.Next()) {
        std::string utt = fst_reader.Key();
        if (!feature_reader.HasKey(utt)) {
          KALDI_WARN << "Not decoding utterance " << utt
                     << " because no features available.";
          num_fail++;
          continue;
        }
        const Matrix<BaseFloat> &features = feature_reader.Value(utt);
        if (features.NumRows() == 0) {
          KALDI_WARN << "Zero-length utterance: " << utt;
          num_fail++;
          continue;
        }
        LatticeBiglmFasterDecoder decoder(fst_reader.Value(), config,
                                          &cache_dfst);
        DecodableAmDiagGmmScaled gmm_decodable(am_gmm, trans_model, features,
                                               acoustic_scale);
        double like;
        if (DecodeUtterance(decoder, gmm_decodable, trans_model, word_syms, utt,
                            acoustic_scale, determinize, allow_partial,
                            &alignment_writer, &words_writer,
                            &compact_lattice_writer, &lattice_writer,
                            &like)) {
          tot_like += like;
          frame_count += features.NumRows();
          num_success++;
        } else num_fail++;
      }
    }
      
    double elapsed = timer.Elapsed();
    KALDI_LOG << "Time taken "<< 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;
    if (num_success != 0) return 0;
    else return 1;
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}