add fill zero opt for frontend

speechx/speechx/decoder/param.h

@@ -20,6 +20,7 @@
 
 // feature
 DEFINE_bool(use_fbank, false, "False for fbank; or linear feature");
+DEFINE_bool(fill_zero, false, "fill zero at last chunk, when chunk < chunk_size");
 // DEFINE_bool(to_float32, true, "audio convert to pcm32. True for linear
 // feature, or fbank");
 DEFINE_int32(num_bins, 161, "num bins of mel");

speechx/speechx/frontend/audio/assembler.cc

@@ -47,17 +47,16 @@ bool Assembler::Read(kaldi::Vector<kaldi::BaseFloat>* feats) {
 // read frame by frame from base_feature_extractor_ into cache_
 bool Assembler::Compute(Vector<BaseFloat>* feats) {
     // compute and feed frame by frame
-    bool result = false;
     while (feature_cache_.size() < frame_chunk_size_) {
         Vector<BaseFloat> feature;
-        result = base_extractor_->Read(&feature);
+        bool result = base_extractor_->Read(&feature);
         if (result == false || feature.Dim() == 0) {
             VLOG(1) << "result: " << result << " feature dim: " << feature.Dim();
             if (IsFinished() == false) {
-                LOG(INFO) << "finished reading feature. cache size: " << feature_cache_.size();
+                VLOG(1) << "finished reading feature. cache size: " << feature_cache_.size();
                 return false;
             } else {
-                LOG(INFO) << "break";
+                VLOG(1) << "break";
                 break;
             }
         }

speechx/speechx/frontend/audio/feature_pipeline.h

@@ -27,6 +27,7 @@
 
 // feature
 DECLARE_bool(use_fbank);
+DECLARE_bool(fill_zero);
 DECLARE_int32(num_bins);
 DECLARE_string(cmvn_file);
 
@@ -80,15 +81,18 @@ struct FeaturePipelineOptions {
 
         // assembler opts
         opts.assembler_opts.subsampling_rate = FLAGS_subsampling_rate;
-        LOG(INFO) << "subsampling rate: "
-                  << opts.assembler_opts.subsampling_rate;
         opts.assembler_opts.receptive_filed_length =
             FLAGS_receptive_field_length;
+        opts.assembler_opts.nnet_decoder_chunk = FLAGS_nnet_decoder_chunk;
+        opts.assembler_opts.fill_zero = FLAGS_fill_zero;
+        LOG(INFO) << "subsampling rate: "
+                  << opts.assembler_opts.subsampling_rate;
         LOG(INFO) << "nnet receptive filed length: "
                   << opts.assembler_opts.receptive_filed_length;
-        opts.assembler_opts.nnet_decoder_chunk = FLAGS_nnet_decoder_chunk;
         LOG(INFO) << "nnet chunk size: "
                   << opts.assembler_opts.nnet_decoder_chunk;
+        LOG(INFO) << "frontend fill zeros: "
+                  << opts.assembler_opts.fill_zero;
         return opts;
     }
 };

speechx/speechx/nnet/decodable.cc

@@ -114,7 +114,7 @@ bool Decodable::AdvanceChunk(kaldi::Vector<kaldi::BaseFloat>* logprobs,
 // read one frame likelihood
 bool Decodable::FrameLikelihood(int32 frame, vector<BaseFloat>* likelihood) {
     if (EnsureFrameHaveComputed(frame) == false) {
-        LOG(INFO) << "framelikehood exit.";
+        VLOG(1) << "framelikehood exit.";
         return false;
     }
 

speechx/speechx/recognizer/recognizer.h

@@ -38,6 +38,8 @@ struct RecognizerResource {
         resource.acoustic_scale = FLAGS_acoustic_scale;
         resource.feature_pipeline_opts =
             FeaturePipelineOptions::InitFromFlags();
+        resource.feature_pipeline_opts.assembler_opts.fill_zero = true;
+        LOG(INFO) << "ds2 need fill zero be true: " << resource.feature_pipeline_opts.assembler_opts.fill_zero;
         resource.model_opts = ModelOptions::InitFromFlags();
         resource.tlg_opts = TLGDecoderOptions::InitFromFlags();
         return resource;

speechx/speechx/recognizer/u2_recognizer.h

@@ -101,6 +101,8 @@ struct U2RecognizerResource {
 
         resource.feature_pipeline_opts =
             ppspeech::FeaturePipelineOptions::InitFromFlags();
+        resource.feature_pipeline_opts.assembler_opts.fill_zero = false;
+        LOG(INFO) << "u2 need fill zero be false: " << resource.feature_pipeline_opts.assembler_opts.fill_zero;
         resource.model_opts = ppspeech::ModelOptions::InitFromFlags();
         resource.decoder_opts = ppspeech::DecodeOptions::InitFromFlags();
         return resource;

speechx/speechx/recognizer/u2_recognizer_main.cc

@@ -85,9 +85,6 @@ int main(int argc, char* argv[]) {
             cnt++;
         }
         CHECK(sample_offset == tot_samples);
-        VLOG(1) << "num decode: " << cnt;
-
-        // recognizer.SetFinished();
 
         // second pass decoding
         recognizer.Rescoring();