[src] Fix to nnet3-chain-get-egs corner-case bug discovered by Armando Muscarellio (kaldi-asr#2536)

danpovey · web-flow · commit 6edf1be704e0 · 2018-07-06T14:23:10.000-04:00
diff --git a/src/chainbin/nnet3-chain-get-egs.cc b/src/chainbin/nnet3-chain-get-egs.cc
@@ -116,6 +116,16 @@ static bool ProcessFile(const TransitionModel *trans_mdl,
                                   supervision_length_tolerance))
     return false;  // LengthsMatch() will have printed a warning.
 
+  // It can happen if people mess with the feature frame-width options, that
+  // there can be small mismatches in length between the supervisions (derived
+  // from lattices) and the features; if this happens, and
+  // supervision_length_tolerance is nonzero, and the num-input-frames is larger
+  // than plausible for this num_output_frames, then it could lead us to try to
+  // access frames in the supervision that don't exist.  The following
+  // if-statement is to prevent that happening.
+  if (num_input_frames > num_output_frames * frame_subsampling_factor)
+    num_input_frames = num_output_frames * frame_subsampling_factor;
+
   std::vector<ChunkTimeInfo> chunks;
 
   utt_splitter->GetChunksForUtterance(num_input_frames, &chunks);
