gmm-est-fmllr-raw.cc File Reference

#include "base/kaldi-common.h"
#include "transform/fmllr-raw.h"
#include "gmm/am-diag-gmm.h"
#include "hmm/transition-model.h"
#include "util/common-utils.h"
#include "hmm/posterior.h"

Include dependency graph for gmm-est-fmllr-raw.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
void	AccStatsForUtterance (const TransitionModel &trans_model, const AmDiagGmm &am_gmm, const Posterior &post, const Matrix< BaseFloat > &feats, FmllrRawAccs *accs)
int	main (int argc, char *argv[])

Function Documentation

main()

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

Definition at line 51 of file gmm-est-fmllr-raw.cc.

References kaldi::AccStatsForUtterance(), count, AmDiagGmm::Dim(), SequentialTableReader< Holder >::Done(), ParseOptions::GetArg(), RandomAccessTableReader< Holder >::HasKey(), rnnlm::i, KALDI_LOG, KALDI_WARN, SequentialTableReader< Holder >::Key(), SequentialTableReader< Holder >::Next(), ParseOptions::NumArgs(), MatrixBase< Real >::NumRows(), ParseOptions::PrintUsage(), AmDiagGmm::Read(), ParseOptions::Read(), TransitionModel::Read(), kaldi::ReadKaldiObject(), FmllrRawOptions::Register(), ParseOptions::Register(), MatrixBase< Real >::SetUnit(), Input::Stream(), RandomAccessTableReader< Holder >::Value(), SequentialTableReader< Holder >::Value(), and TableWriter< Holder >::Write().

                                 {
  try {
    typedef kaldi::int32 int32;
    using namespace kaldi;
    const char *usage =
        "Estimate fMLLR transforms in the space before splicing and linear transforms\n"
        "such as LDA+MLLT, but using models in the space transformed by these transforms\n"
        "Requires the original spliced features, and the full LDA+MLLT (or similar) matrix\n"
        "including the 'rejected' rows (see the program get-full-lda-mat)\n"
        "Usage: gmm-est-fmllr-raw [options] <model-in> <full-lda-mat-in> "
        "<feature-rspecifier> <post-rspecifier> <transform-wspecifier>\n";


    int32 raw_feat_dim = 13;
    ParseOptions po(usage);
    FmllrRawOptions opts;
    std::string spk2utt_rspecifier;
    po.Register("spk2utt", &spk2utt_rspecifier, "rspecifier for speaker to "
                "utterance-list map");
    po.Register("raw-feat-dim", &raw_feat_dim, "Dimension of raw features "
                "prior to splicing");
    opts.Register(&po);

    po.Read(argc, argv);

    if (po.NumArgs() != 5) {
      po.PrintUsage();
      exit(1);
    }
    
    std::string model_rxfilename = po.GetArg(1),
        full_lda_mat_rxfilename = po.GetArg(2),
        feature_rspecifier = po.GetArg(3),
        post_rspecifier = po.GetArg(4),
        transform_wspecifier = po.GetArg(5);

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

    Matrix<BaseFloat> full_lda_mat;
    ReadKaldiObject(full_lda_mat_rxfilename, &full_lda_mat);
    
    RandomAccessPosteriorReader post_reader(post_rspecifier);
    BaseFloatMatrixWriter transform_writer(transform_wspecifier);
    
    double tot_auxf_impr = 0.0, tot_count = 0.0;
    
    int32 num_done = 0, num_err = 0;
    if (!spk2utt_rspecifier.empty()) { // Adapting per speaker
      SequentialTokenVectorReader spk2utt_reader(spk2utt_rspecifier);
      RandomAccessBaseFloatMatrixReader feature_reader(feature_rspecifier);
      
      for (; !spk2utt_reader.Done(); spk2utt_reader.Next()) {
        FmllrRawAccs accs(raw_feat_dim, am_gmm.Dim(), full_lda_mat);
        std::string spk = spk2utt_reader.Key();
        const std::vector<std::string> &uttlist = spk2utt_reader.Value();
        for (size_t i = 0; i < uttlist.size(); i++) {
          std::string utt = uttlist[i];
          if (!feature_reader.HasKey(utt)) {
            KALDI_WARN << "Features not found for utterance " << utt;
            num_err++;
            continue;
          }
          if (!post_reader.HasKey(utt)) {
            KALDI_WARN << "Posteriors not found for utterance " << utt;
            num_err++;
            continue;
          }
          const Matrix<BaseFloat> &feats = feature_reader.Value(utt);
          const Posterior &post = post_reader.Value(utt);
          if (static_cast<int32>(post.size()) != feats.NumRows()) {
            KALDI_WARN << "Size mismatch between posteriors " << post.size()
                       << " and features " << feats.NumRows();
            num_err++;
            continue;
          }

          AccStatsForUtterance(trans_model, am_gmm, post, feats, &accs);
          num_done++;
        }
        
        BaseFloat auxf_impr, count;
        {
          Matrix<BaseFloat> transform(raw_feat_dim, raw_feat_dim + 1);
          transform.SetUnit();
          accs.Update(opts, &transform, &auxf_impr, &count);
          transform_writer.Write(spk, transform);
        }
        KALDI_LOG << "For speaker " << spk << ", auxf-impr from raw fMLLR is "
                  << (auxf_impr/count) << " over " << count << " frames.";
        tot_auxf_impr += auxf_impr;
        tot_count += count;
      }
    } else {  // --spk2utt option not given -> adapt per utterance.
      SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
      for (; !feature_reader.Done(); feature_reader.Next()) {
        std::string utt = feature_reader.Key();
        if (!post_reader.HasKey(utt)) {
          KALDI_WARN << "Posteriors not found for utterance " << utt;
          num_err++;
          continue;
        }
        const Matrix<BaseFloat> &feats = feature_reader.Value();
        const Posterior &post = post_reader.Value(utt);

        if (static_cast<int32>(post.size()) != feats.NumRows()) {
          KALDI_WARN << "Size mismatch between posteriors " << post.size()
                     << " and features " << feats.NumRows();
          num_err++;
          continue;
        }

        FmllrRawAccs accs(raw_feat_dim, am_gmm.Dim(), full_lda_mat);

        AccStatsForUtterance(trans_model, am_gmm, post, feats, &accs);
        
        BaseFloat auxf_impr, count;        
        {
          Matrix<BaseFloat> transform(raw_feat_dim, raw_feat_dim + 1);
          transform.SetUnit();
          accs.Update(opts, &transform, &auxf_impr, &count);
          transform_writer.Write(utt, transform);
        }
        KALDI_LOG << "For utterance " << utt << ", auxf-impr from raw fMLLR is "
                  << (auxf_impr/count) << " over " << count << " frames.";
        tot_auxf_impr += auxf_impr;
        tot_count += count;
        num_done++;
      }
    }

    KALDI_LOG << "Processed " << num_done << " utterances, "
              << num_err << " had errors.";
    KALDI_LOG << "Overall raw-fMLLR auxf impr per frame is "
              << (tot_auxf_impr / tot_count) << " over " << tot_count
              << " frames.";
    return (num_done != 0 ? 0 : 1);
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}