Fixes OnlineFeInput, which was broken by commit cc2469e.

We have reverted just the parts affecting OnlineFeInput.

Since ExtractWaveformRemainder was removed in commit 1180e46, we recovered and adapted that code and used it directly in OnlineFeInput<E>::Compute.

We have also created a fake FrameExtractionOptions so that NumFrames() will still work as expected.

Note that we also support the snip_edges=false case (OnlineFeInput constructor now takes an additional optional parameter).

Author: jdieguez

src/online/online-feat-input.h

@@ -31,6 +31,7 @@
 
 #include "online-audio-source.h"
 #include "feat/feature-functions.h"
+#include "feat/feature-window.h"
 
 namespace kaldi {
 
@@ -275,7 +276,8 @@ class OnlineFeInput : public OnlineFeatInputItf {
   // "frame_size" - frame extraction window size in audio samples
   // "frame_shift" - feature frame width in audio samples
   OnlineFeInput(OnlineAudioSourceItf *au_src, E *fe,
-                const int32 frame_size, const int32 frame_shift);
+                const int32 frame_size, const int32 frame_shift,
+                const bool snip_edges = true);
 
   virtual int32 Dim() const { return extractor_->Dim(); }
 
@@ -287,15 +289,26 @@ class OnlineFeInput : public OnlineFeatInputItf {
   const int32 frame_size_;
   const int32 frame_shift_;
   Vector<BaseFloat> wave_; // the samples to be passed for extraction
+  Vector<BaseFloat> wave_remainder_; // the samples remained from the previous
+                                     // feature batch
+  FrameExtractionOptions frame_opts_;
 
   KALDI_DISALLOW_COPY_AND_ASSIGN(OnlineFeInput);
 };
 
 template<class E>
 OnlineFeInput<E>::OnlineFeInput(OnlineAudioSourceItf *au_src, E *fe,
-                                   int32 frame_size, int32 frame_shift)
+                                int32 frame_size, int32 frame_shift,
+                                bool snip_edges)
     : source_(au_src), extractor_(fe),
-      frame_size_(frame_size), frame_shift_(frame_shift) {}
+      frame_size_(frame_size), frame_shift_(frame_shift) {
+      // we need a FrameExtractionOptions to call NumFrames()
+      // 1000 is just a fake sample rate which equates ms and samples
+      frame_opts_.samp_freq = 1000;
+      frame_opts_.frame_shift_ms = frame_shift;
+      frame_opts_.frame_length_ms = frame_size;
+      frame_opts_.snip_edges = snip_edges;
+}
 
 template<class E> bool
 OnlineFeInput<E>::Compute(Matrix<BaseFloat> *output) {
@@ -311,11 +324,26 @@ OnlineFeInput<E>::Compute(Matrix<BaseFloat> *output) {
 
   bool ans = source_->Read(&read_samples);
 
+  Vector<BaseFloat> all_samples(wave_remainder_.Dim() + read_samples.Dim());
+  all_samples.Range(0, wave_remainder_.Dim()).CopyFromVec(wave_remainder_);
+  all_samples.Range(wave_remainder_.Dim(), read_samples.Dim()).
+      CopyFromVec(read_samples);
+
   // Extract the features
-  if (read_samples.Dim() >= frame_size_) {
-    extractor_->Compute(read_samples, 1.0, output);
+  if (all_samples.Dim() >= frame_size_) {
+    // extract waveform remainder before calling Compute()
+    int32 num_frames = NumFrames(all_samples.Dim(), frame_opts_);
+    // offset is the amount at the start that has been extracted.
+    int32 offset = num_frames * frame_shift_;
+    int32 remaining_len = all_samples.Dim() - offset;
+    wave_remainder_.Resize(remaining_len);
+    KALDI_ASSERT(remaining_len >= 0);
+    if (remaining_len > 0)
+      wave_remainder_.CopyFromVec(SubVector<BaseFloat>(all_samples, offset, remaining_len));
+    extractor_->Compute(all_samples, 1.0, output);
   } else {
     output->Resize(0, 0);
+    wave_remainder_ = all_samples;
   }
 
   return ans;