SingleUtteranceGmmDecoder Class Reference

You will instantiate this class when you want to decode a single utterance using the online-decoding setup. More...

#include <online-gmm-decoding.h>

Collaboration diagram for SingleUtteranceGmmDecoder:

Public Member Functions
	SingleUtteranceGmmDecoder (const OnlineGmmDecodingConfig &config, const OnlineGmmDecodingModels &models, const OnlineFeaturePipeline &feature_prototype, const fst::Fst< fst::StdArc > &fst, const OnlineGmmAdaptationState &adaptation_state)
OnlineFeaturePipeline &	FeaturePipeline ()
void	AdvanceDecoding ()
	advance the decoding as far as we can. More...
void	FinalizeDecoding ()
	Finalize the decoding. More...
bool	HaveTransform () const
	Returns true if we already have an fMLLR transform. More...
void	EstimateFmllr (bool end_of_utterance)
	Estimate the [basis-]fMLLR transform and apply it to the features. More...
void	GetAdaptationState (OnlineGmmAdaptationState *adaptation_state) const
void	GetLattice (bool rescore_if_needed, bool end_of_utterance, CompactLattice *clat) const
	Gets the lattice. More...
void	GetBestPath (bool end_of_utterance, Lattice *best_path) const
	Outputs an FST corresponding to the single best path through the current lattice. More...
BaseFloat	FinalRelativeCost ()
	This function outputs to "final_relative_cost", if non-NULL, a number >= 0 that will be close to zero if the final-probs were close to the best probs active on the final frame. More...
bool	EndpointDetected (const OnlineEndpointConfig &config)
	This function calls EndpointDetected from online-endpoint.h, with the required arguments. More...
	~SingleUtteranceGmmDecoder ()

Private Member Functions
bool	GetGaussianPosteriors (bool end_of_utterance, GaussPost *gpost)
bool	RescoringIsNeeded () const
	Returns true if doing a lattice rescoring pass would have any point, i.e. More...

Private Attributes
OnlineGmmDecodingConfig	config_
std::vector< int32 >	silence_phones_
const OnlineGmmDecodingModels &	models_
OnlineFeaturePipeline *	feature_pipeline_
const OnlineGmmAdaptationState &	orig_adaptation_state_
OnlineGmmAdaptationState	adaptation_state_
LatticeFasterOnlineDecoder	decoder_

Detailed Description

You will instantiate this class when you want to decode a single utterance using the online-decoding setup.

This is an alternative to manually putting things together yourself.

Definition at line 216 of file online-gmm-decoding.h.

Constructor & Destructor Documentation

SingleUtteranceGmmDecoder()

SingleUtteranceGmmDecoder	(	const OnlineGmmDecodingConfig &	config,
		const OnlineGmmDecodingModels &	models,
		const OnlineFeaturePipeline &	feature_prototype,
		const fst::Fst< fst::StdArc > &	fst,
		const OnlineGmmAdaptationState &	adaptation_state
	)

Definition at line 48 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::adaptation_state_, SingleUtteranceGmmDecoder::config_, SingleUtteranceGmmDecoder::decoder_, SingleUtteranceGmmDecoder::feature_pipeline_, LatticeFasterDecoderTpl< FST, Token >::InitDecoding(), KALDI_ERR, OnlineFeaturePipeline::SetTransform(), OnlineGmmDecodingConfig::silence_phones, SingleUtteranceGmmDecoder::silence_phones_, kaldi::SortAndUniq(), kaldi::SplitStringToIntegers(), and OnlineGmmAdaptationState::transform.

                                                     :
    config_(config), models_(models),
    feature_pipeline_(feature_prototype.New()),
    orig_adaptation_state_(adaptation_state),
    adaptation_state_(adaptation_state),
    decoder_(fst, config.faster_decoder_opts) {
  if (!SplitStringToIntegers(config_.silence_phones, ":", false,
                             &silence_phones_))
    KALDI_ERR << "Bad --silence-phones option '"
              << config_.silence_phones << "'";
  SortAndUniq(&silence_phones_);
  feature_pipeline_->SetTransform(adaptation_state_.transform);
  decoder_.InitDecoding();
}

~SingleUtteranceGmmDecoder()

~SingleUtteranceGmmDecoder

(

)

Definition at line 305 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::feature_pipeline_.

                                                      {
  delete feature_pipeline_;
}

Member Function Documentation

AdvanceDecoding()

void AdvanceDecoding

(

)

advance the decoding as far as we can.

May also estimate fMLLR after advancing the decoding, depending on the configuration values in config_.adaptation_policy_opts. [Note: we expect the user will also call EstimateFmllr() at utterance end, which should generally improve the quality of the estimated transforms, although we don't rely on this].

Definition at line 69 of file online-gmm-decoding.cc.

References OnlineGmmDecodingConfig::acoustic_scale, OnlineGmmDecodingConfig::adaptation_policy_opts, LatticeFasterDecoderTpl< FST, Token >::AdvanceDecoding(), SingleUtteranceGmmDecoder::config_, SingleUtteranceGmmDecoder::decoder_, OnlineGmmDecodingAdaptationPolicyConfig::DoAdapt(), SingleUtteranceGmmDecoder::EstimateFmllr(), SingleUtteranceGmmDecoder::feature_pipeline_, OnlineFeaturePipeline::FrameShiftInSeconds(), OnlineGmmDecodingModels::GetModel(), OnlineGmmDecodingModels::GetOnlineAlignmentModel(), OnlineGmmDecodingModels::GetTransitionModel(), SingleUtteranceGmmDecoder::HaveTransform(), SingleUtteranceGmmDecoder::models_, LatticeFasterDecoderTpl< FST, Token >::NumFramesDecoded(), MatrixBase< Real >::NumRows(), SingleUtteranceGmmDecoder::orig_adaptation_state_, and OnlineGmmAdaptationState::transform.

                                                {

  const AmDiagGmm &am_gmm = (HaveTransform() ? models_.GetModel() :
                             models_.GetOnlineAlignmentModel());

  // The decodable object is lightweight, we lose nothing
  // from constructing it each time we want to decode more of the
  // input.
  DecodableDiagGmmScaledOnline decodable(am_gmm,
                                         models_.GetTransitionModel(),
                                         config_.acoustic_scale,
                                         feature_pipeline_);

  int32 old_frames = decoder_.NumFramesDecoded();
  
  // This will decode as many frames as are currently available.
  decoder_.AdvanceDecoding(&decodable);

  
  {  // possibly estimate fMLLR.
    int32 new_frames = decoder_.NumFramesDecoded();
    BaseFloat frame_shift = feature_pipeline_->FrameShiftInSeconds();
    // if the original adaptation state (at utterance-start) had no transform,
    // then this means it's the first utt of the speaker... even if not, if we
    // don't have a transform it probably makes sense to treat it as the 1st utt
    // of the speaker, i.e. to do fMLLR adaptation sooner.
    bool is_first_utterance_of_speaker =
        (orig_adaptation_state_.transform.NumRows() == 0);
    bool end_of_utterance = false;
    if (config_.adaptation_policy_opts.DoAdapt(old_frames * frame_shift,
                                               new_frames * frame_shift,
                                               is_first_utterance_of_speaker))
      this->EstimateFmllr(end_of_utterance);
  }
}

EndpointDetected()

bool EndpointDetected

(

const OnlineEndpointConfig &

config

)

This function calls EndpointDetected from online-endpoint.h, with the required arguments.

Definition at line 310 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::decoder_, kaldi::EndpointDetected(), SingleUtteranceGmmDecoder::feature_pipeline_, OnlineFeaturePipeline::FrameShiftInSeconds(), OnlineGmmDecodingModels::GetTransitionModel(), and SingleUtteranceGmmDecoder::models_.

                                        {
  const TransitionModel &tmodel = models_.GetTransitionModel();
  return kaldi::EndpointDetected(config, tmodel,
                                 feature_pipeline_->FrameShiftInSeconds(),
                                 decoder_);
}

EstimateFmllr()

void EstimateFmllr

(

bool

end_of_utterance

)

Estimate the [basis-]fMLLR transform and apply it to the features.

This will get used if you call RescoreLattice() or if you just continue decoding; however to get it applied retroactively you'd have to call RescoreLattice(). "end_of_utterance" just affects how we interpret the final-probs in the lattice. This should generally be true if you think you've reached the end of the grammar, and false otherwise.

Definition at line 208 of file online-gmm-decoding.cc.

References FmllrDiagGmmAccs::AccumulateFromPosteriors(), SingleUtteranceGmmDecoder::adaptation_state_, OnlineGmmDecodingConfig::basis_opts, AffineXformStats::beta_, BasisFmllrEstimate::ComputeTransform(), SingleUtteranceGmmDecoder::config_, SingleUtteranceGmmDecoder::decoder_, AffineXformStats::Dim(), VectorBase< Real >::Dim(), BasisFmllrEstimate::Dim(), OnlineFeaturePipeline::Dim(), SingleUtteranceGmmDecoder::feature_pipeline_, OnlineFeaturePipeline::FreezeCmvn(), OnlineFeaturePipeline::GetAsMatrix(), OnlineGmmDecodingModels::GetFmllrBasis(), OnlineFeaturePipeline::GetFrame(), SingleUtteranceGmmDecoder::GetGaussianPosteriors(), OnlineGmmDecodingModels::GetModel(), AmDiagGmm::GetPdf(), kaldi::GetVerboseLevel(), rnnlm::i, FmllrDiagGmmAccs::Init(), rnnlm::j, KALDI_ERR, KALDI_VLOG, KALDI_WARN, SingleUtteranceGmmDecoder::models_, LatticeFasterDecoderTpl< FST, Token >::NumFramesDecoded(), MatrixBase< Real >::NumRows(), SingleUtteranceGmmDecoder::orig_adaptation_state_, OnlineFeaturePipeline::SetTransform(), OnlineGmmAdaptationState::spk_stats, and OnlineGmmAdaptationState::transform.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding().

                                                                   {
  if (decoder_.NumFramesDecoded() == 0) {
    KALDI_WARN << "You have decoded no data so cannot estimate fMLLR.";
  }

  if (GetVerboseLevel() >= 2) {
    Matrix<BaseFloat> feats;
    feature_pipeline_->GetAsMatrix(&feats);
    KALDI_VLOG(2) << "Features are " << feats;
  }
  

  GaussPost gpost;
  GetGaussianPosteriors(end_of_utterance, &gpost);
  
  FmllrDiagGmmAccs &spk_stats = adaptation_state_.spk_stats;
  
  if (spk_stats.beta_ !=
      orig_adaptation_state_.spk_stats.beta_) {
    // This could happen if the user called EstimateFmllr() twice on the
    // same utterance... we don't want to count any stats twice so we
    // have to reset the stats to what they were before this utterance
    // (possibly empty).
    spk_stats = orig_adaptation_state_.spk_stats;
  }
  
  int32 dim = feature_pipeline_->Dim();
  if (spk_stats.Dim() == 0)
    spk_stats.Init(dim);
  
  Matrix<BaseFloat> empty_transform;
  feature_pipeline_->SetTransform(empty_transform);
  Vector<BaseFloat> feat(dim);

  if (adaptation_state_.transform.NumRows() == 0) {
    // If this is the first time we're estimating fMLLR, freeze the CMVN to its
    // current value.  It doesn't matter too much what value this is, since we
    // have already computed the Gaussian-level alignments (it may have a small
    // effect if the basis is very small and doesn't include an offset as part
    // of the transform).
    feature_pipeline_->FreezeCmvn();
  }
  
  // GetModel() returns the model to be used for estimating
  // transforms.
  const AmDiagGmm &am_gmm = models_.GetModel();
  
  for (size_t i = 0; i < gpost.size(); i++) {
    feature_pipeline_->GetFrame(i, &feat);    
    for (size_t j = 0; j < gpost[i].size(); j++) {
      int32 pdf_id = gpost[i][j].first; // caution: this gpost has pdf-id
                                        // instead of transition-id, which is
                                        // unusual.
      const Vector<BaseFloat> &posterior(gpost[i][j].second);
      spk_stats.AccumulateFromPosteriors(am_gmm.GetPdf(pdf_id),
                                         feat, posterior);
    }
  }
  
  const BasisFmllrEstimate &basis = models_.GetFmllrBasis();
  if (basis.Dim() == 0)
    KALDI_ERR << "In order to estimate fMLLR, you need to supply the "
              << "--fmllr-basis option.";
  Vector<BaseFloat> basis_coeffs;
  BaseFloat impr = basis.ComputeTransform(spk_stats,
                                          &adaptation_state_.transform,
                                          &basis_coeffs, config_.basis_opts);
  KALDI_VLOG(3) << "Objective function improvement from basis-fMLLR is "
                << (impr / spk_stats.beta_) << " per frame, over "
                << spk_stats.beta_ << " frames, #params estimated is "
                << basis_coeffs.Dim();
  feature_pipeline_->SetTransform(adaptation_state_.transform);
}

FeaturePipeline()

OnlineFeaturePipeline& FeaturePipeline ( )

inline

Definition at line 224 of file online-gmm-decoding.h.

{ return *feature_pipeline_; }

FinalizeDecoding()

void FinalizeDecoding

(

)

Finalize the decoding.

Cleanups and prunes remaining tokens, so the final result is faster to obtain.

Definition at line 105 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::decoder_, and LatticeFasterDecoderTpl< FST, Token >::FinalizeDecoding().

                                                 {
  decoder_.FinalizeDecoding();
}

FinalRelativeCost()

BaseFloat FinalRelativeCost ( )

inline

This function outputs to "final_relative_cost", if non-NULL, a number >= 0 that will be close to zero if the final-probs were close to the best probs active on the final frame.

(the output to final_relative_cost is based on the first-pass decoding). If it's close to zero (e.g. < 5, as a guess), it means you reached the end of the grammar with good probability, which can be taken as a good sign that the input was OK.

Definition at line 276 of file online-gmm-decoding.h.

References kaldi::EndpointDetected().

{ return decoder_.FinalRelativeCost(); }

GetAdaptationState()

void GetAdaptationState

(

OnlineGmmAdaptationState *

adaptation_state

)

const

Definition at line 287 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::adaptation_state_, OnlineGmmAdaptationState::cmvn_state, SingleUtteranceGmmDecoder::feature_pipeline_, and OnlineFeaturePipeline::GetCmvnState().

                                                      {
  *adaptation_state = adaptation_state_;
  feature_pipeline_->GetCmvnState(&adaptation_state->cmvn_state);
}

GetBestPath()

void GetBestPath	(	bool	end_of_utterance,
		Lattice *	best_path
	)		const

Outputs an FST corresponding to the single best path through the current lattice.

If "use_final_probs" is true AND we reached the final-state of the graph then it will include those as final-probs, else it will treat all final-probs as one.

Definition at line 341 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::decoder_, and LatticeFasterOnlineDecoderTpl< FST >::GetBestPath().

                                                                      {
  decoder_.GetBestPath(best_path, end_of_utterance);
}

GetGaussianPosteriors()

bool GetGaussianPosteriors ( bool  end_of_utterance, GaussPost *  gpost  )

private

Definition at line 111 of file online-gmm-decoding.cc.

References DiagGmm::ComponentPosteriors(), SingleUtteranceGmmDecoder::config_, kaldi::ConvertPosteriorToPdfs(), SingleUtteranceGmmDecoder::decoder_, fst::DeterminizeLatticePruned(), OnlineFeaturePipeline::Dim(), SingleUtteranceGmmDecoder::feature_pipeline_, OnlineGmmDecodingConfig::fmllr_lattice_beam, OnlineFeaturePipeline::GetFrame(), OnlineGmmDecodingModels::GetModel(), OnlineGmmDecodingModels::GetOnlineAlignmentModel(), AmDiagGmm::GetPdf(), LatticeFasterOnlineDecoderTpl< FST >::GetRawLatticePruned(), OnlineGmmDecodingModels::GetTransitionModel(), SingleUtteranceGmmDecoder::HaveTransform(), rnnlm::i, rnnlm::j, KALDI_ASSERT, KALDI_VLOG, KALDI_WARN, kaldi::LatticeForwardBackward(), SingleUtteranceGmmDecoder::models_, LatticeFasterDecoderTpl< FST, Token >::NumFramesDecoded(), kaldi::PruneLattice(), VectorBase< Real >::Scale(), SingleUtteranceGmmDecoder::silence_phones_, OnlineGmmDecodingConfig::silence_weight, kaldi::TopSortLatticeIfNeeded(), and kaldi::WeightSilencePost().

Referenced by SingleUtteranceGmmDecoder::EstimateFmllr().

                                                                        {
  // Gets the Gaussian-level posteriors for this utterance, using whatever
  // features and model we are currently decoding with.  We'll use these
  // to estimate basis-fMLLR with.
  if (decoder_.NumFramesDecoded() == 0) {
    KALDI_WARN << "You have decoded no data so cannot estimate fMLLR.";
    return false;
  }
  
  KALDI_ASSERT(config_.fmllr_lattice_beam > 0.0);
  
  // Note: we'll just use whatever acoustic scaling factor we were decoding
  // with.  This is in the lattice that we get from decoder_.GetRawLattice().
  Lattice raw_lat;
  decoder_.GetRawLatticePruned(&raw_lat, end_of_utterance,
                               config_.fmllr_lattice_beam);
  
  // At this point we could rescore the lattice if we wanted, and
  // this might improve the accuracy on long utterances that were
  // the first utterance of that speaker, if we had already
  // estimated the fMLLR by the time we reach this code (e.g. this
  // was the second call).  We don't do this right now.
  
  PruneLattice(config_.fmllr_lattice_beam, &raw_lat);

#if 1 // Do determinization. 
  Lattice det_lat; // lattice-determinized lattice-- represent this as Lattice
                   // not CompactLattice, as LatticeForwardBackward() does not
                   // accept CompactLattice.


  fst::Invert(&raw_lat); // want to determinize on words.
  fst::ILabelCompare<kaldi::LatticeArc> ilabel_comp;
  fst::ArcSort(&raw_lat, ilabel_comp); // improves efficiency of determinization
  
  fst::DeterminizeLatticePruned(raw_lat,
                                double(config_.fmllr_lattice_beam),
                                &det_lat);

  fst::Invert(&det_lat); // invert back.
  
  if (det_lat.NumStates() == 0) {
    // Do nothing if the lattice is empty.  This should not happen.
    KALDI_WARN << "Got empty lattice.  Not estimating fMLLR.";
    return false;
  }
#else
  Lattice &det_lat = raw_lat; // Don't determinize.
#endif
  TopSortLatticeIfNeeded(&det_lat);
  
  // Note: the acoustic scale we use here is whatever we decoded with.
  Posterior post;
  BaseFloat tot_fb_like = LatticeForwardBackward(det_lat, &post);

  KALDI_VLOG(3) << "Lattice forward-backward likelihood was "
                << (tot_fb_like / post.size()) << " per frame over " << post.size()
                << " frames.";

  ConstIntegerSet<int32> silence_set(silence_phones_);  // faster lookup
  const TransitionModel &trans_model = models_.GetTransitionModel();
  WeightSilencePost(trans_model, silence_set,
                    config_.silence_weight, &post);  
  
  const AmDiagGmm &am_gmm = (HaveTransform() ? models_.GetModel() :
                             models_.GetOnlineAlignmentModel());


  Posterior pdf_post;
  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
  
  Vector<BaseFloat> feat(feature_pipeline_->Dim());

  double tot_like = 0.0, tot_weight = 0.0;
  gpost->resize(pdf_post.size());
  for (size_t i = 0; i < pdf_post.size(); i++) {
    feature_pipeline_->GetFrame(i, &feat);
    for (size_t j = 0; j < pdf_post[i].size(); j++) {
      int32 pdf_id = pdf_post[i][j].first;
      BaseFloat weight = pdf_post[i][j].second;
      const DiagGmm &gmm = am_gmm.GetPdf(pdf_id);
      Vector<BaseFloat> this_post_vec;
      BaseFloat like = gmm.ComponentPosteriors(feat, &this_post_vec);
      this_post_vec.Scale(weight);
      tot_like += like * weight;
      tot_weight += weight;
      (*gpost)[i].push_back(std::make_pair(pdf_id, this_post_vec));
    }
  }
  KALDI_VLOG(3) << "Average likelihood weighted by posterior was "
                << (tot_like / tot_weight) << " over " << tot_weight
                << " frames (after downweighting silence).";  
  return true;
}

GetLattice()

void GetLattice	(	bool	rescore_if_needed,
		bool	end_of_utterance,
		CompactLattice *	clat
	)		const

Gets the lattice.

If rescore_if_needed is true, and if there is any point in rescoring the state-level lattice (see RescoringIsNeeded()), it will rescore the lattice. The output lattice has any acoustic scaling in it (which will typically be desirable in an online-decoding context); if you want an un-scaled lattice, scale it using ScaleLattice() with the inverse of the acoustic weight. "end_of_utterance" will be true if you want the final-probs to be included.

Definition at line 318 of file online-gmm-decoding.cc.

References OnlineGmmDecodingConfig::acoustic_scale, SingleUtteranceGmmDecoder::config_, SingleUtteranceGmmDecoder::decoder_, LatticeFasterDecoderConfig::det_opts, fst::DeterminizeLatticePhonePrunedWrapper(), OnlineGmmDecodingConfig::faster_decoder_opts, SingleUtteranceGmmDecoder::feature_pipeline_, OnlineGmmDecodingModels::GetFinalModel(), LatticeFasterDecoderTpl< FST, Token >::GetRawLattice(), OnlineGmmDecodingModels::GetTransitionModel(), KALDI_WARN, LatticeFasterDecoderConfig::lattice_beam, SingleUtteranceGmmDecoder::models_, kaldi::PruneLattice(), kaldi::RescoreLattice(), and SingleUtteranceGmmDecoder::RescoringIsNeeded().

                                                                       {
  Lattice lat;
  double lat_beam = config_.faster_decoder_opts.lattice_beam;
  decoder_.GetRawLattice(&lat, end_of_utterance);
  if (rescore_if_needed && RescoringIsNeeded()) {
    DecodableDiagGmmScaledOnline decodable(models_.GetFinalModel(),
                                           models_.GetTransitionModel(),
                                           config_.acoustic_scale,
                                           feature_pipeline_);

    if (!kaldi::RescoreLattice(&decodable, &lat))
      KALDI_WARN << "Error rescoring lattice";
  }
  PruneLattice(lat_beam, &lat);

  DeterminizeLatticePhonePrunedWrapper(models_.GetTransitionModel(),
                                       &lat, lat_beam, clat,
                                       config_.faster_decoder_opts.det_opts);
  
}

HaveTransform()

bool HaveTransform

(

)

const

Returns true if we already have an fMLLR transform.

The user will already know this; the call is for convenience.

Definition at line 283 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::feature_pipeline_, and OnlineFeaturePipeline::HaveFmllrTransform().

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), and SingleUtteranceGmmDecoder::GetGaussianPosteriors().

                                                    {
  return (feature_pipeline_->HaveFmllrTransform());
}

RescoringIsNeeded()

bool RescoringIsNeeded ( ) const

private

Returns true if doing a lattice rescoring pass would have any point, i.e.

if we have estimated fMLLR during this utterance, or if we have a discriminative model that differs from the fMLLR model *and* we currently have fMLLR features.

Definition at line 293 of file online-gmm-decoding.cc.

References SingleUtteranceGmmDecoder::adaptation_state_, MatrixBase< Real >::ApproxEqual(), OnlineGmmDecodingModels::GetFinalModel(), OnlineGmmDecodingModels::GetModel(), SingleUtteranceGmmDecoder::models_, MatrixBase< Real >::NumRows(), SingleUtteranceGmmDecoder::orig_adaptation_state_, and OnlineGmmAdaptationState::transform.

Referenced by SingleUtteranceGmmDecoder::GetLattice().

                                                        {
  if (orig_adaptation_state_.transform.NumRows() !=
      adaptation_state_.transform.NumRows()) return true;  // fMLLR was estimated
  if (!orig_adaptation_state_.transform.ApproxEqual(
          adaptation_state_.transform)) return true;  // fMLLR was re-estimated
  if (adaptation_state_.transform.NumRows() != 0 &&
      &models_.GetModel() != &models_.GetFinalModel())
    return true; // we have an fMLLR transform, and a discriminatively estimated
                 // model which differs from the one used to estimate fMLLR.
  return false;
}

Member Data Documentation

adaptation_state_

OnlineGmmAdaptationState adaptation_state_

private

Definition at line 303 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::EstimateFmllr(), SingleUtteranceGmmDecoder::GetAdaptationState(), SingleUtteranceGmmDecoder::RescoringIsNeeded(), and SingleUtteranceGmmDecoder::SingleUtteranceGmmDecoder().

config_

OnlineGmmDecodingConfig config_

private

Definition at line 293 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), SingleUtteranceGmmDecoder::EstimateFmllr(), SingleUtteranceGmmDecoder::GetGaussianPosteriors(), SingleUtteranceGmmDecoder::GetLattice(), and SingleUtteranceGmmDecoder::SingleUtteranceGmmDecoder().

decoder_

LatticeFasterOnlineDecoder decoder_

private

Definition at line 304 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), SingleUtteranceGmmDecoder::EndpointDetected(), SingleUtteranceGmmDecoder::EstimateFmllr(), SingleUtteranceGmmDecoder::FinalizeDecoding(), SingleUtteranceGmmDecoder::GetBestPath(), SingleUtteranceGmmDecoder::GetGaussianPosteriors(), SingleUtteranceGmmDecoder::GetLattice(), and SingleUtteranceGmmDecoder::SingleUtteranceGmmDecoder().

feature_pipeline_

OnlineFeaturePipeline* feature_pipeline_

private

Definition at line 297 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), SingleUtteranceGmmDecoder::EndpointDetected(), SingleUtteranceGmmDecoder::EstimateFmllr(), SingleUtteranceGmmDecoder::GetAdaptationState(), SingleUtteranceGmmDecoder::GetGaussianPosteriors(), SingleUtteranceGmmDecoder::GetLattice(), SingleUtteranceGmmDecoder::HaveTransform(), SingleUtteranceGmmDecoder::SingleUtteranceGmmDecoder(), and SingleUtteranceGmmDecoder::~SingleUtteranceGmmDecoder().

models_

const OnlineGmmDecodingModels& models_

private

Definition at line 296 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), SingleUtteranceGmmDecoder::EndpointDetected(), SingleUtteranceGmmDecoder::EstimateFmllr(), SingleUtteranceGmmDecoder::GetGaussianPosteriors(), SingleUtteranceGmmDecoder::GetLattice(), and SingleUtteranceGmmDecoder::RescoringIsNeeded().

orig_adaptation_state_

const OnlineGmmAdaptationState& orig_adaptation_state_

private

Definition at line 298 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::AdvanceDecoding(), SingleUtteranceGmmDecoder::EstimateFmllr(), and SingleUtteranceGmmDecoder::RescoringIsNeeded().

silence_phones_

std::vector<int32> silence_phones_

private

Definition at line 294 of file online-gmm-decoding.h.

Referenced by SingleUtteranceGmmDecoder::GetGaussianPosteriors(), and SingleUtteranceGmmDecoder::SingleUtteranceGmmDecoder().

---

The documentation for this class was generated from the following files:

• online2/online-gmm-decoding.h
• online2/online-gmm-decoding.cc