Decoding works, results are empty yet

python/example/batch/test_batch.py

@@ -1,21 +1,24 @@
 #!/usr/bin/env python3
 
-from vosk import Model, BatchRecognizer
+from vosk import Model, BatchRecognizer, GpuInit, GpuThreadInit
 import sys
 import os
 import wave
 
+GpuInit()
+GpuThreadInit()
+
 model = Model("model")
 rec = BatchRecognizer(model, 16000.0)
 
 fnames = open("tedlium.list").readlines()
-fds = [open(x) for x in fnames]
+fds = [open(x.strip(), "rb") for x in fnames]
 ended = set()
 while True:
-    for i, fd in fds:
-        if i in ended():
+    for i, fd in enumerate(fds):
+        if i in ended:
             continue
-        data = fd.read(4000)
+        data = fd.read(16000)
         if len(data) == 0:
             rec.FinishStream(i)
             ended.add(i)

python/vosk/__init__.py

@@ -117,7 +117,7 @@ def AcceptWaveform(self, uid, data):
         res = _c.vosk_batch_recognizer_accept_waveform(self._handle, uid, data, len(data))
 
     def Results(self):
-        return _ffi.string(_c.vosk_batch_recognizer_result(self._handle)).decode('utf-8')
+        return _ffi.string(_c.vosk_batch_recognizer_results(self._handle)).decode('utf-8')
 
     def FinishStream(self, uid):
-        _c.vosk_recognizer_final_result(self._handle, uid)
+        _c.vosk_batch_recognizer_finish_stream(self._handle, uid)

src/Makefile

@@ -37,17 +37,19 @@ CFLAGS=-g -O3 -std=c++17 -Wno-deprecated-declarations -fPIC -DFST_NO_DYNAMIC_LIN
 	-I. -I$(KALDI_ROOT)/src -I$(OPENFST_ROOT)/include $(EXTRA_CFLAGS) 
 
 LIBS= \
+        $(KALDI_ROOT)/src/cudadecoder/kaldi-cudadecoder.a \
+        $(KALDI_ROOT)/src/cudafeat/kaldi-cudafeat.a \
 	$(KALDI_ROOT)/src/online2/kaldi-online2.a \
 	$(KALDI_ROOT)/src/decoder/kaldi-decoder.a \
 	$(KALDI_ROOT)/src/ivector/kaldi-ivector.a \
 	$(KALDI_ROOT)/src/gmm/kaldi-gmm.a \
-	$(KALDI_ROOT)/src/nnet3/kaldi-nnet3.a \
 	$(KALDI_ROOT)/src/tree/kaldi-tree.a \
 	$(KALDI_ROOT)/src/feat/kaldi-feat.a \
 	$(KALDI_ROOT)/src/lat/kaldi-lat.a \
 	$(KALDI_ROOT)/src/lm/kaldi-lm.a \
 	$(KALDI_ROOT)/src/rnnlm/kaldi-rnnlm.a \
 	$(KALDI_ROOT)/src/hmm/kaldi-hmm.a \
+	$(KALDI_ROOT)/src/nnet3/kaldi-nnet3.a \
 	$(KALDI_ROOT)/src/transform/kaldi-transform.a \
 	$(KALDI_ROOT)/src/cudamatrix/kaldi-cudamatrix.a \
 	$(KALDI_ROOT)/src/matrix/kaldi-matrix.a \
@@ -68,7 +70,7 @@ ifeq ($(HAVE_OPENBLAS_CLAPACK), 1)
 endif
 
 ifeq ($(HAVE_MKL), 1)
-    CFLAGS += -I$(MKL_ROOT)/include
+    CFLAGS += -DHAVE_MKL=1 -I$(MKL_ROOT)/include
     LIBS += -L$(MKL_ROOT)/lib/intel64 -Wl,-rpath=$(MKL_ROOT)/lib/intel64 -lmkl_rt -lmkl_intel_lp64 -lmkl_core -lmkl_sequential
 endif
 
@@ -79,8 +81,6 @@ endif
 ifeq ($(HAVE_CUDA), 1)
     CFLAGS+=-DHAVE_CUDA=1 -I$(CUDA_ROOT)/include
     LIBS+=\
-        $(KALDI_ROOT)/src/cudadecoder/kaldi-cudadecoder.a \
-        $(KALDI_ROOT)/src/cudafeat/kaldi-cudafeat.a \
         -L$(CUDA_ROOT)/lib64 -lcuda -lcublas -lcusparse -lcudart -lcurand -lcufft -lcusolver -lnvToolsExt
 endif
 

src/batch_recognizer.cc

@@ -25,9 +25,22 @@ BatchRecognizer::BatchRecognizer(Model *model, float sample_frequency) : model_(
     model_->Ref();
 
     BatchedThreadedNnet3CudaOnlinePipelineConfig batched_decoder_config;
+    batched_decoder_config.num_worker_threads = 4;
+    batched_decoder_config.max_batch_size = 100;
+
+    batched_decoder_config.feature_opts.feature_type = "mfcc";
+    batched_decoder_config.feature_opts.mfcc_config = "model/conf/mfcc.conf";
+    batched_decoder_config.feature_opts.ivector_extraction_config = "model/conf/ivector.conf";
+    batched_decoder_config.decoder_opts.max_active = 7000;
+    batched_decoder_config.decoder_opts.default_beam = 13.0;
+    batched_decoder_config.decoder_opts.lattice_beam = 8.0;
+    batched_decoder_config.compute_opts.acoustic_scale = 1.0;
+    batched_decoder_config.compute_opts.frame_subsampling_factor = 3;
+    batched_decoder_config.compute_opts.frames_per_chunk = 312;
 
     cuda_pipeline_ = new BatchedThreadedNnet3CudaOnlinePipeline 
          (batched_decoder_config, *model_->hclg_fst_, *model_->nnet_, *model_->trans_model_);
+    cuda_pipeline_->SetSymbolTable(*model_->word_syms_);
 
     CudaOnlinePipelineDynamicBatcherConfig dynamic_batcher_config;
     dynamic_batcher_ = new CudaOnlinePipelineDynamicBatcher(dynamic_batcher_config,
@@ -60,6 +73,9 @@ void BatchRecognizer::InitRescoring()
 
 void BatchRecognizer::FinishStream(uint64_t id)
 {
+    Vector<BaseFloat> wave;
+    SubVector<BaseFloat> chunk(wave.Data(), 0);
+    dynamic_batcher_->Push(id, false, true, chunk);
     streams_.erase(id);
 }
 
@@ -77,7 +93,7 @@ void BatchRecognizer::AcceptWaveform(uint64_t id, const char *data, int len)
           [&, id](const std::string &str, bool partial,
                        bool endpoint_detected) {
               if (partial) {
-                  KALDI_LOG << "id #" << id << " [partial] : " << str;
+                  KALDI_LOG << "id #" << id << " [partial] : " << str << ":";
               }
 
               if (endpoint_detected) {

src/model.cc

@@ -241,9 +241,9 @@ void Model::ReadDataFiles()
         SetDropoutTestMode(true, &(nnet_->GetNnet()));
         nnet3::CollapseModel(nnet3::CollapseModelConfig(), &(nnet_->GetNnet()));
     }
-    decodable_info_ = new nnet3::DecodableNnetSimpleLoopedInfo(decodable_opts_,
-                                                               nnet_);
 
+/*    decodable_info_ = new nnet3::DecodableNnetSimpleLoopedInfo(decodable_opts_,
+                                                               nnet_);
     if (stat(final_ie_rxfilename_.c_str(), &buffer) == 0) {
         KALDI_LOG << "Loading i-vector extractor from " << final_ie_rxfilename_;
 
@@ -261,7 +261,7 @@ void Model::ReadDataFiles()
     } else {
         feature_info_.use_ivectors = false;
     }
-
+*/
     if (stat(global_cmvn_stats_rxfilename_.c_str(), &buffer) == 0) {
         KALDI_LOG << "Reading CMVN stats from " << global_cmvn_stats_rxfilename_;
         feature_info_.use_cmvn = true;

src/vosk_api.cc

@@ -193,12 +193,12 @@ void vosk_batch_recognizer_free(VoskBatchRecognizer *recognizer)
     delete ((BatchRecognizer *)recognizer);
 }
 
-void vosk_batch_recognizer_accept_waveform(VoskRecognizer *recognizer, int id, const char *data, int length)
+void vosk_batch_recognizer_accept_waveform(VoskBatchRecognizer *recognizer, int id, const char *data, int length)
 {
     ((BatchRecognizer *)recognizer)->AcceptWaveform(id, data, length);
 }
 
-void vosk_batch_recognizer_finish_stream(VoskRecognizer *recognizer, int id)
+void vosk_batch_recognizer_finish_stream(VoskBatchRecognizer *recognizer, int id)
 {
     ((BatchRecognizer *)recognizer)->FinishStream(id);
 }

src/vosk_api.h

@@ -301,10 +301,10 @@ VoskBatchRecognizer *vosk_batch_recognizer_new(VoskModel *model, float sample_fr
 void vosk_batch_recognizer_free(VoskBatchRecognizer *recognizer);
 
 /** Accept batch voice data */
-void vosk_batch_recognizer_accept_waveform(VoskRecognizer *recognizer, int id, const char *data, int length);
+void vosk_batch_recognizer_accept_waveform(VoskBatchRecognizer *recognizer, int id, const char *data, int length);
 
 /** Closes the stream */
-void vosk_batch_recognizer_finish_stream(VoskRecognizer *recognizer, int id);
+void vosk_batch_recognizer_finish_stream(VoskBatchRecognizer *recognizer, int id);
 
 /** Return results */
 const char *vosk_batch_recognizer_results(VoskBatchRecognizer *recognizer);