These are the stats for a particular utterance, i.e. More...

#include <ivector-extractor.h>

Collaboration diagram for IvectorExtractorUtteranceStats:

[legend]

Public Member Functions
	IvectorExtractorUtteranceStats (int32 num_gauss, int32 feat_dim, bool need_2nd_order_stats)

void	AccStats (const MatrixBase< BaseFloat > &feats, const Posterior &post)

void	Scale (double scale)

double	NumFrames ()

Protected Attributes
Vector< double >	gamma_

Matrix< double >	X_

std::vector< SpMatrix< double > >	S_

Friends
class	IvectorExtractor

class	IvectorExtractorStats

Detailed Description

These are the stats for a particular utterance, i.e.

the sufficient stats for estimating an iVector (if need_2nd_order_stats == true, we can also estimate the variance of the model; these 2nd order stats are not needed if we just need the iVector.

Definition at line 79 of file ivector-extractor.h.

Constructor & Destructor Documentation

◆ IvectorExtractorUtteranceStats()

IvectorExtractorUtteranceStats	(	int32	num_gauss,
		int32	feat_dim,
		bool	need_2nd_order_stats
	)

inline

Definition at line 81 of file ivector-extractor.h.

References rnnlm::i.

                                                            :
       gamma_(num_gauss), X_(num_gauss, feat_dim) {
     if (need_2nd_order_stats) {
       S_.resize(num_gauss);
       for (int32 i = 0; i < num_gauss; i++)
         S_[i].Resize(feat_dim);
     }
   }

Member Function Documentation

◆ AccStats()

void AccStats	(	const MatrixBase< BaseFloat > &	feats,
		const Posterior &	post
	)

Definition at line 852 of file ivector-extractor.cc.

References SpMatrix< Real >::AddVec2(), rnnlm::i, KALDI_ASSERT, MatrixBase< Real >::NumCols(), MatrixBase< Real >::NumRows(), and PackedMatrix< Real >::SetZero().

Referenced by IvectorExtractorStats::AccStatsForUtterance(), IvectorExtractTask::operator()(), and kaldi::TestIvectorExtraction().

                            {
   typedef std::vector<std::pair<int32, BaseFloat> > VecType;
   int32 num_frames = feats.NumRows(),
       num_gauss = X_.NumRows(),
       feat_dim = feats.NumCols();
   KALDI_ASSERT(X_.NumCols() == feat_dim);
   KALDI_ASSERT(feats.NumRows() == static_cast<int32>(post.size()));
   bool update_variance = (!S_.empty());
   SpMatrix<double> outer_prod(feat_dim);
   for (int32 t = 0; t < num_frames; t++) {
     SubVector<BaseFloat> frame(feats, t);
     const VecType &this_post(post[t]);
     if (update_variance) {
       outer_prod.SetZero();
       outer_prod.AddVec2(1.0, frame);
     }
     for (VecType::const_iterator iter = this_post.begin();
          iter != this_post.end(); ++iter) {
       int32 i = iter->first; // Gaussian index.
       KALDI_ASSERT(i >= 0 && i < num_gauss &&
                    "Out-of-range Gaussian (mismatched posteriors?)");
       double weight = iter->second;
       gamma_(i) += weight;
       X_.Row(i).AddVec(weight, frame);
       if (update_variance)
         S_[i].AddSp(weight, outer_prod);
     }
   }
 }

◆ NumFrames()

double NumFrames ( )

inline

Definition at line 96 of file ivector-extractor.h.

Referenced by kaldi::TestIvectorExtraction().

96 { return gamma_.Sum(); }

kaldi::IvectorExtractorUtteranceStats::gamma_

Vector< double > gamma_

Definition: ivector-extractor.h:101

kaldi::VectorBase::Sum

Real Sum() const

Returns sum of the elements.

Definition: kaldi-vector.cc:688

◆ Scale()

void Scale ( double scale )

Definition at line 884 of file ivector-extractor.cc.

References rnnlm::i, and OnlineIvectorEstimationStats::Scale().

                                                        {
   gamma_.Scale(scale);
   X_.Scale(scale);
   for (size_t i = 0; i < S_.size(); i++)
     S_[i].Scale(scale);
 }