Fix nnet3 endpointing to correctly use frame subsampling factor (kaldi-asr#1184)

Tanel Alumäe · Tanel Alumäe · commit 465b73cb178e · 2016-11-15T23:31:30.000+02:00
diff --git a/src/nnet3/online-nnet3-decodable-simple.h b/src/nnet3/online-nnet3-decodable-simple.h
@@ -102,6 +102,7 @@ class DecodableNnet3SimpleOnline: public DecodableInterface {
   /// Indices are one-based!  This is for compatibility with OpenFst.
   virtual int32 NumIndices() const { return trans_model_.NumTransitionIds(); }
 
+  virtual int32 FrameSubsamplingFactor() const { return opts_.frame_subsampling_factor; }
  private:
 
   /// If the neural-network outputs for this frame are not cached, it computes
diff --git a/src/online2/online-nnet3-decoding.cc b/src/online2/online-nnet3-decoding.cc
@@ -72,8 +72,9 @@ void SingleUtteranceNnet3Decoder::GetBestPath(bool end_of_utterance,
 
 bool SingleUtteranceNnet3Decoder::EndpointDetected(
     const OnlineEndpointConfig &config) {
+  int32 subsample = decodable_.FrameSubsamplingFactor();
   return kaldi::EndpointDetected(config, tmodel_,
-                                 feature_pipeline_->FrameShiftInSeconds(),
+                                 feature_pipeline_->FrameShiftInSeconds() * subsample,
                                  decoder_);  
 }
 

Original file line number	Diff line number	Diff line change
`@@ -72,8 +72,9 @@ void SingleUtteranceNnet3Decoder::GetBestPath(bool end_of_utterance,`
`72`	`72`
`73`	`73`	`bool SingleUtteranceNnet3Decoder::EndpointDetected(`
`74`	`74`	`const OnlineEndpointConfig &config) {`
	`75`	`+ int32 subsample = decodable_.FrameSubsamplingFactor();`
`75`	`76`	`return kaldi::EndpointDetected(config, tmodel_,`
`76`		`- feature_pipeline_->FrameShiftInSeconds(),`
	`77`	`+ feature_pipeline_->FrameShiftInSeconds() * subsample,`
`77`	`78`	`decoder_);`
`78`	`79`	`}`
`79`	`80`