[engine] add wfst recognizer in example (#3173)

* update wfst script

* add skip blank

runtime/engine/asr/decoder/ctc_prefix_beam_search_decoder.cc

@@ -87,9 +87,9 @@ void CTCPrefixBeamSearch::AdvanceDecode(
 
         VLOG(1) << "num_frame_decoded_: " << num_frame_decoded_;
     }
-    VLOG(1) << "AdvanceDecode feat + forward  cost: " << feat_nnet_cost
+    VLOG(2) << "AdvanceDecode feat + forward  cost: " << feat_nnet_cost
             << " sec.";
-    VLOG(1) << "AdvanceDecode search  cost: " << search_cost << " sec.";
+    VLOG(2) << "AdvanceDecode search  cost: " << search_cost << " sec.";
 }
 
 static bool PrefixScoreCompare(

runtime/engine/asr/decoder/ctc_prefix_beam_search_decoder_main.cc

@@ -71,7 +71,7 @@ int main(int argc, char* argv[]) {
     std::shared_ptr<ppspeech::DataCache> raw_data =
         std::make_shared<ppspeech::DataCache>();
     std::shared_ptr<ppspeech::NnetProducer> nnet_producer =
-        std::make_shared<ppspeech::NnetProducer>(nnet, raw_data);
+        std::make_shared<ppspeech::NnetProducer>(nnet, raw_data, 1.0);
     std::shared_ptr<ppspeech::Decodable> decodable =
         std::make_shared<ppspeech::Decodable>(nnet_producer);
 

runtime/engine/asr/decoder/ctc_tlg_decoder.h

@@ -44,7 +44,7 @@ struct TLGDecoderOptions {
         decoder_opts.word_symbol_table = FLAGS_word_symbol_table;
         decoder_opts.fst_path = FLAGS_graph_path;
         LOG(INFO) << "fst path: " << decoder_opts.fst_path;
-        LOG(INFO) << "fst symbole table: " << decoder_opts.word_symbol_table;
+        LOG(INFO) << "symbole table: " << decoder_opts.word_symbol_table;
 
         if (!decoder_opts.fst_path.empty()) {
             CHECK(FileExists(decoder_opts.fst_path));

runtime/engine/asr/decoder/ctc_tlg_decoder_main.cc

@@ -54,7 +54,7 @@ int main(int argc, char* argv[]) {
     ppspeech::ModelOptions model_opts = ppspeech::ModelOptions::InitFromFlags();
 
     std::shared_ptr<ppspeech::NnetProducer> nnet_producer =
-        std::make_shared<ppspeech::NnetProducer>(nullptr);
+        std::make_shared<ppspeech::NnetProducer>(nullptr, nullptr, 1.0);
     std::shared_ptr<ppspeech::Decodable> decodable(
         new ppspeech::Decodable(nnet_producer, FLAGS_acoustic_scale));
 

runtime/engine/asr/decoder/param.h

@@ -35,13 +35,11 @@ DEFINE_int32(subsampling_rate,
              "two CNN(kernel=3) module downsampling rate.");
 DEFINE_int32(nnet_decoder_chunk, 1, "paddle nnet forward chunk");
 
-
 // nnet
 DEFINE_string(model_path, "avg_1.jit.pdmodel", "paddle nnet model");
 #ifdef USE_ONNX
 DEFINE_bool(with_onnx_model, false, "True mean the model path is onnx model path");
 #endif
-//DEFINE_string(param_path, "avg_1.jit.pdiparams", "paddle nnet model param");
 
 // decoder
 DEFINE_double(acoustic_scale, 1.0, "acoustic scale");
@@ -50,10 +48,9 @@ DEFINE_string(word_symbol_table, "", "word symbol table");
 DEFINE_int32(max_active, 7500, "max active");
 DEFINE_double(beam, 15.0, "decoder beam");
 DEFINE_double(lattice_beam, 7.5, "decoder beam");
-
+DEFINE_double(blank_threshold, 0.98, "blank skip threshold");
 
 // DecodeOptions flags
-// DEFINE_int32(chunk_size, -1, "decoding chunk size");
 DEFINE_int32(num_left_chunks, -1, "left chunks in decoding");
 DEFINE_double(ctc_weight,
               0.5,

runtime/engine/asr/nnet/nnet_producer.cc

@@ -22,8 +22,9 @@ using kaldi::BaseFloat;
 using std::vector;
 
 NnetProducer::NnetProducer(std::shared_ptr<NnetBase> nnet,
-                           std::shared_ptr<FrontendInterface> frontend)
-    : nnet_(nnet), frontend_(frontend) {
+                           std::shared_ptr<FrontendInterface> frontend,
+                           float blank_threshold)
+    : nnet_(nnet), frontend_(frontend), blank_threshold_(blank_threshold) {
     Reset();
 }
 
@@ -45,7 +46,6 @@ void NnetProducer::Acceptlikelihood(
 
 bool NnetProducer::Read(std::vector<kaldi::BaseFloat>* nnet_prob) {
     bool flag = cache_.pop(nnet_prob);
-    VLOG(1) << "nnet cache_ size: " << cache_.size();
     return flag;
 }
 
@@ -70,7 +70,22 @@ bool NnetProducer::Compute() {
         std::vector<BaseFloat> logprob(
             out.logprobs.data() + idx * vocab_dim,
             out.logprobs.data() + (idx + 1) * vocab_dim);
-        cache_.push_back(logprob);
+        // process blank prob
+        float blank_prob = std::exp(logprob[0]);
+        if (blank_prob > blank_threshold_) {
+            last_frame_logprob_ = logprob;
+            is_last_frame_skip_ = true;
+            continue;
+        } else {
+            int cur_max = std::max(logprob.begin(), logprob.end()) - logprob.begin();
+            if (cur_max == last_max_elem_ && cur_max != 0 && is_last_frame_skip_) {
+                cache_.push_back(last_frame_logprob_);
+                last_max_elem_ = cur_max;
+            }
+            last_max_elem_ = cur_max;
+            is_last_frame_skip_ = false; 
+            cache_.push_back(logprob);
+        }
     }
     return true;
 }

runtime/engine/asr/nnet/nnet_producer.h

@@ -24,7 +24,8 @@ namespace ppspeech {
 class NnetProducer {
   public:
     explicit NnetProducer(std::shared_ptr<NnetBase> nnet,
-                          std::shared_ptr<FrontendInterface> frontend = NULL);
+                          std::shared_ptr<FrontendInterface> frontend,
+                          float blank_threshold);
     // Feed feats or waves
     void Accept(const std::vector<kaldi::BaseFloat>& inputs);
 
@@ -64,6 +65,10 @@ class NnetProducer {
     std::shared_ptr<FrontendInterface> frontend_;
     std::shared_ptr<NnetBase> nnet_;
     SafeQueue<std::vector<kaldi::BaseFloat>> cache_;
+    std::vector<BaseFloat> last_frame_logprob_;
+    bool is_last_frame_skip_ = false;
+    int last_max_elem_ = -1;
+    float blank_threshold_ = 0.0;
     bool finished_;
 
     DISALLOW_COPY_AND_ASSIGN(NnetProducer);

runtime/engine/asr/nnet/u2_nnet.cc

@@ -124,7 +124,15 @@ U2Nnet::U2Nnet(const U2Nnet& other) {
     offset_ = other.offset_;
 
     // copy model ptr
-    model_ = other.model_->Clone();
+    // model_ = other.model_->Clone();
+    // hack, fix later
+    #ifdef WITH_GPU
+        dev_ = phi::GPUPlace();
+    #else
+        dev_ = phi::CPUPlace();
+    #endif
+    paddle::jit::Layer model = paddle::jit::Load(other.opts_.model_path, dev_);
+    model_ = std::make_shared<paddle::jit::Layer>(std::move(model));
     ctc_activation_ = model_->Function("ctc_activation");
     subsampling_rate_ = model_->Attribute<int>("subsampling_rate");
     right_context_ = model_->Attribute<int>("right_context");
@@ -166,6 +174,7 @@ void U2Nnet::Reset() {
         std::move(paddle::zeros({0, 0, 0, 0}, paddle::DataType::FLOAT32));
 
     encoder_outs_.clear();
+    VLOG(1) << "FeedForward cost: " << cost_time_ << " sec. ";
     VLOG(3) << "u2nnet reset";
 }
 
@@ -185,8 +194,10 @@ void U2Nnet::FeedForward(const std::vector<BaseFloat>& features,
     std::vector<kaldi::BaseFloat> ctc_probs;
     ForwardEncoderChunkImpl(
         features, feature_dim, &out->logprobs, &out->vocab_dim);
-    VLOG(1) << "FeedForward cost: " << timer.Elapsed() << " sec. "
+    float forward_chunk_time = timer.Elapsed();
+    VLOG(1) << "FeedForward cost: " << forward_chunk_time << " sec. "
             << features.size() / feature_dim << " frames.";
+    cost_time_ += forward_chunk_time;
 }
 
 

runtime/engine/asr/nnet/u2_nnet.h

@@ -113,8 +113,8 @@ class U2Nnet : public U2NnetBase {
     void EncoderOuts(
         std::vector<std::vector<kaldi::BaseFloat>>* encoder_out) const;
 
+    ModelOptions opts_; // hack, fix later
   private:
-    ModelOptions opts_;
 
     phi::Place dev_;
     std::shared_ptr<paddle::jit::Layer> model_{nullptr};
@@ -127,6 +127,7 @@ class U2Nnet : public U2NnetBase {
     paddle::jit::Function forward_encoder_chunk_;
     paddle::jit::Function forward_attention_decoder_;
     paddle::jit::Function ctc_activation_;
+    float cost_time_ = 0.0;
 };
 
 }  // namespace ppspeech
\ No newline at end of file

runtime/engine/asr/recognizer/recognizer_controller_impl.cc

@@ -21,6 +21,7 @@ namespace ppspeech {
 RecognizerControllerImpl::RecognizerControllerImpl(const RecognizerResource& resource)
 : opts_(resource) {
     BaseFloat am_scale = resource.acoustic_scale;
+    BaseFloat blank_threshold = resource.blank_threshold;
     const FeaturePipelineOptions& feature_opts = resource.feature_pipeline_opts;
     std::shared_ptr<FeaturePipeline> feature_pipeline(
         new FeaturePipeline(feature_opts));
@@ -34,7 +35,7 @@ RecognizerControllerImpl::RecognizerControllerImpl(const RecognizerResource& res
         nnet = resource.nnet->Clone();
     }
 #endif
-    nnet_producer_.reset(new NnetProducer(nnet, feature_pipeline));
+    nnet_producer_.reset(new NnetProducer(nnet, feature_pipeline, blank_threshold));
     nnet_thread_ = std::thread(RunNnetEvaluation, this);
 
     decodable_.reset(new Decodable(nnet_producer_, am_scale));

runtime/engine/asr/recognizer/recognizer_main.cc

@@ -88,7 +88,8 @@ int main(int argc, char* argv[]) {
 
         kaldi::Timer timer;
         recognizer_ptr->AttentionRescoring();
-        tot_attention_rescore_time += timer.Elapsed();
+        float rescore_time = timer.Elapsed();
+        tot_attention_rescore_time += rescore_time;
 
         std::string result = recognizer_ptr->GetFinalResult();
         if (result.empty()) {
@@ -101,7 +102,7 @@ int main(int argc, char* argv[]) {
         tot_decode_time += local_timer.Elapsed();
         LOG(INFO) << utt << " " << result;
         LOG(INFO) << " RTF: " << local_timer.Elapsed() / dur << " dur: " << dur
-                  << " cost: " << local_timer.Elapsed();
+                  << " cost: " << local_timer.Elapsed() << " rescore:" << rescore_time;
 
         result_writer.Write(utt, result);
 

runtime/engine/asr/recognizer/recognizer_resource.h

@@ -12,6 +12,7 @@ DECLARE_double(reverse_weight);
 DECLARE_int32(nbest);
 DECLARE_int32(blank);
 DECLARE_double(acoustic_scale);
+DECLARE_double(blank_threshold);
 DECLARE_string(word_symbol_table);
 
 namespace ppspeech {
@@ -71,6 +72,7 @@ struct DecodeOptions {
 struct RecognizerResource {
     // decodable opt 
     kaldi::BaseFloat acoustic_scale{1.0};
+    kaldi::BaseFloat blank_threshold{0.98};
 
     FeaturePipelineOptions feature_pipeline_opts{};
     ModelOptions model_opts{};
@@ -80,6 +82,7 @@ struct RecognizerResource {
     static RecognizerResource InitFromFlags() {
         RecognizerResource resource;
         resource.acoustic_scale = FLAGS_acoustic_scale;
+        resource.blank_threshold = FLAGS_blank_threshold;
         LOG(INFO) << "acoustic_scale: " << resource.acoustic_scale;
 
         resource.feature_pipeline_opts =

runtime/engine/kaldi/fstbin/CMakeLists.txt

@@ -11,5 +11,5 @@ fsttablecompose
 foreach(binary IN LISTS BINS)
    add_executable(${binary} ${CMAKE_CURRENT_SOURCE_DIR}/${binary}.cc) 
    target_include_directories(${binary} PRIVATE ${SPEECHX_ROOT} ${SPEECHX_ROOT}/kaldi)
-   target_link_libraries(${binary} PUBLIC kaldi-fstext glog libgflags_nothreads.so fst dl)
+   target_link_libraries(${binary} PUBLIC kaldi-fstext glog gflags fst dl)
 endforeach()

runtime/examples/u2pp_ol/wenetspeech/RESULTS.md

@@ -4,7 +4,7 @@
 
 ## U2++ Attention Rescore
 
-> Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz, support `avx512_vnni`
+> Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, support `avx512_vnni`
 > RTF with feature and decoder which is more end to end.
 
 ### FP32
@@ -23,18 +23,15 @@ Other -> 100.00 % N=3 C=0 S=3 D=0 I=0
 #### RTF 
 
 ```
-I1027 10:52:38.662868 51665 u2_recognizer_main.cc:122] total wav duration is: 36108.9 sec
-I1027 10:52:38.662858 51665 u2_recognizer_main.cc:121] total cost:11169.1 sec
-I1027 10:52:38.662876 51665 u2_recognizer_main.cc:123] RTF is: 0.309318
+I1027 10:52:38.662868 51665 recognizer_main.cc:122] total wav duration is: 36108.9 sec
+I1027 10:52:38.662858 51665 recognizer_main.cc:121] total cost:9577.31 sec
+I1027 10:52:38.662876 51665 recognizer_main.cc:123] RTF is: 0.265234
 ```
 
 ### INT8
 
 `local/recognizer_quant.sh`
 
-> RTF relative improve 12.8%, which count feature and decoder time.
-> Test under Paddle commit c331e2ce2031d68a553bc9469a07c30d718438f3  
-
 #### CER
 
 ```
@@ -52,16 +49,22 @@ I1110 09:59:52.551717 37249 u2_recognizer_main.cc:123] total decode cost:9737.63
 I1110 09:59:52.551723 37249 u2_recognizer_main.cc:124] RTF is: 0.269674
 ```
 
-### CTC Prefix Beam Search
+### TLG decoder without attention rescore
 
-`local/decode.sh`
+`local/recognizer_wfst.sh`
 
 #### CER
 
 ```
-Overall -> 6.74 % N=104765 C=98106 S=6516 D=143 I=401
-Mandarin -> 6.74 % N=104762 C=98106 S=6513 D=143 I=401
-English -> 0.00 % N=0 C=0 S=0 D=0 I=0
+Overall -> 4.73 % N=104765 C=100001 S=4283 D=481 I=187
+Mandarin -> 4.72 % N=104762 C=100001 S=4280 D=481 I=187
 Other -> 100.00 % N=3 C=0 S=3 D=0 I=0
+```
 
+#### RTF
+```
+I0417 08:07:15.300631 75784 recognizer_main.cc:113] total wav duration is: 36108.9 sec
+I0417 08:07:15.300642 75784 recognizer_main.cc:114] total decode cost:10247.7 sec
+I0417 08:07:15.300648 75784 recognizer_main.cc:115] total rescore cost:908.228 sec
+I0417 08:07:15.300653 75784 recognizer_main.cc:116] RTF is: 0.283
 ```

runtime/examples/u2pp_ol/wenetspeech/local/recognizer_fastdeploy.sh

+#!/bin/bash
+set -e
+
+data=data
+exp=exp
+nj=20
+
+. utils/parse_options.sh
+
+mkdir -p $exp
+ckpt_dir=./data/model
+model_dir=$ckpt_dir/onnx_model/
+aishell_wav_scp=aishell_test.scp
+text=$data/test/text
+
+./local/split_data.sh $data $data/$aishell_wav_scp $aishell_wav_scp $nj
+
+utils/run.pl JOB=1:$nj $data/split${nj}/JOB/recognizer.fd.log \
+recognizer_main \
+    --use_fbank=true \
+    --num_bins=80 \
+    --model_path=$model_dir \
+    --word_symbol_table=$model_dir/unit.txt \
+    --nnet_decoder_chunk=16 \
+    --receptive_field_length=7 \
+    --subsampling_rate=4 \
+    --with_onnx_model=true \
+    --wav_rspecifier=scp:$data/split${nj}/JOB/${aishell_wav_scp} \
+    --result_wspecifier=ark,t:$data/split${nj}/JOB/recognizer.fd.rsl.ark
+
+
+cat $data/split${nj}/*/recognizer.fd.rsl.ark > $exp/aishell.recognizer.fd.rsl
+utils/compute-wer.py --char=1 --v=1 $text $exp/aishell.recognizer.fd.rsl > $exp/aishell.recognizer.fd.err
+echo "recognizer fd test have finished!!!"
+echo "please checkout in $exp/aishell.recognizer.fd.err"
+tail -n 7 $exp/aishell.recognizer.fd.err

runtime/examples/u2pp_ol/wenetspeech/local/recognizer_quant.sh

@@ -16,7 +16,7 @@ text=$data/test/text
 ./local/split_data.sh $data $data/$aishell_wav_scp $aishell_wav_scp $nj
 
 utils/run.pl JOB=1:$nj $data/split${nj}/JOB/recognizer.quant.log \
-u2_recognizer_main \
+recognizer_main \
     --use_fbank=true \
     --num_bins=80 \
     --cmvn_file=$model_dir/mean_std.json \

runtime/examples/u2pp_ol/wenetspeech/local/recognizer_wfst.sh

@@ -3,7 +3,7 @@ set -e
 
 data=data
 exp=exp
-nj=40
+nj=20
 
 . utils/parse_options.sh
 
@@ -19,6 +19,15 @@ lang_dir=./data/lang_test/
 graph=$lang_dir/TLG.fst
 word_table=$lang_dir/words.txt
 
+if [ ! -f $graph ]; then
+    # download ngram, if you want to make graph by yourself, please refer local/run_build_tlg.sh
+    mkdir -p $lang_dir
+    pushd $lang_dir
+    wget -c https://paddlespeech.bj.bcebos.com/speechx/examples/ngram/zh/tlg.zip
+    unzip tlg.zip
+    popd
+fi
+
 utils/run.pl JOB=1:$nj $data/split${nj}/JOB/recognizer_wfst.log \
 recognizer_main \
     --use_fbank=true \
@@ -31,6 +40,8 @@ recognizer_main \
     --receptive_field_length=7 \
     --subsampling_rate=4 \
     --wav_rspecifier=scp:$data/split${nj}/JOB/${aishell_wav_scp} \
+    --rescoring_weight=0.0 \
+    --acoustic_scale=2 \
     --result_wspecifier=ark,t:$data/split${nj}/JOB/result_recognizer_wfst.ark
 
 

runtime/examples/u2pp_ol/wenetspeech/local/recognizer_wfst_fastdeploy.sh

+#!/bin/bash
+set -e
+
+data=data
+exp=exp
+nj=20
+
+. utils/parse_options.sh
+
+mkdir -p $exp
+ckpt_dir=./data/model
+model_dir=$ckpt_dir/onnx_model/
+aishell_wav_scp=aishell_test.scp
+text=$data/test/text
+
+./local/split_data.sh $data $data/$aishell_wav_scp $aishell_wav_scp $nj
+
+lang_dir=./data/lang_test/
+graph=$lang_dir/TLG.fst
+word_table=$lang_dir/words.txt
+
+if [ ! -f $graph ]; then
+    # download ngram, if you want to make graph by yourself, please refer local/run_build_tlg.sh
+    mkdir -p $lang_dir
+    pushd $lang_dir
+    wget -c https://paddlespeech.bj.bcebos.com/speechx/examples/ngram/zh/tlg.zip
+    unzip tlg.zip
+    popd
+fi
+
+utils/run.pl JOB=1:$nj $data/split${nj}/JOB/recognizer_wfst_fd.log \
+recognizer_main \
+    --use_fbank=true \
+    --num_bins=80 \
+    --model_path=$model_dir \
+    --graph_path=$lang_dir/TLG.fst \
+    --word_symbol_table=$word_table \
+    --nnet_decoder_chunk=16 \
+    --receptive_field_length=7 \
+    --subsampling_rate=4 \
+    --wav_rspecifier=scp:$data/split${nj}/JOB/${aishell_wav_scp} \
+    --rescoring_weight=0.0 \
+    --acoustic_scale=2 \
+    --result_wspecifier=ark,t:$data/split${nj}/JOB/result_recognizer_wfst_fd.ark
+
+
+cat $data/split${nj}/*/result_recognizer_wfst_fd.ark > $exp/aishell_recognizer_wfst_fd
+utils/compute-wer.py --char=1 --v=1 $text $exp/aishell_recognizer_wfst_fd > $exp/aishell.recognizer_wfst_fd.err
+echo "recognizer test have finished!!!"
+echo "please checkout in $exp/aishell.recognizer_wfst_fd.err"
+tail -n 7 $exp/aishell.recognizer_wfst_fd.err

runtime/examples/u2pp_ol/wenetspeech/local/run_build_tlg.sh

@@ -7,13 +7,12 @@ set -eo pipefail
 # different acustic model has different vocab
 ckpt_dir=data/model/asr1_chunk_conformer_u2pp_wenetspeech_static_1.3.0.model
 unit=$ckpt_dir/vocab.txt       # vocab file, line: char/spm_pice
-model_dir=$ckpt_dir/exp/deepspeech2_online/checkpoints/
 
 stage=2
 stop_stage=100
 corpus=aishell
 lexicon=data/lexicon.txt  # line: word ph0 ... phn, aishell/resource_aishell/lexicon.txt
-text=data/text            # line: utt text, aishell/data_aishell/transcript/aishell_transcript_v0.8.txt
+text=data/text            # line: utt text, aishell/data_aishell/transcript/aishell_transcript_v0.8.txt filter by data/train/text
 
 . utils/parse_options.sh
 

runtime/examples/u2pp_ol/wenetspeech/path.sh

@@ -12,7 +12,7 @@ TOOLS_BIN=$ENGINE_TOOLS/valgrind/install/bin
 
 export LC_AL=C
 
-export PATH=$PATH:$TOOLS_BIN:$ENGINE_BUILD/nnet:$ENGINE_BUILD/decoder:$ENGINE_BUILD/../common/frontend/audio:$ENGINE_BUILD/recognizer
+export PATH=$PATH:$TOOLS_BIN:$ENGINE_BUILD/nnet:$ENGINE_BUILD/decoder:$ENGINE_BUILD/../common/frontend/audio:$ENGINE_BUILD/recognizer:../../../fc_patch/openfst/bin:$ENGINE_BUILD/../kaldi/fstbin:$ENGINE_BUILD/../kaldi/lmbin
 
 #PADDLE_LIB_PATH=$(python -c "import os; import paddle; include_dir=paddle.sysconfig.get_include(); paddle_dir=os.path.split(include_dir)[0]; libs_dir=os.path.join(paddle_dir, 'libs'); fluid_dir=os.path.join(paddle_dir, 'fluid'); out=':'.join([libs_dir, fluid_dir]); print(out);")
 export LD_LIBRARY_PATH=$PADDLE_LIB_PATH:$LD_LIBRARY_PATH

runtime/examples/u2pp_ol/wenetspeech/run.sh

@@ -69,23 +69,17 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ];then
     fi
 fi
 
-
 if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
-    # process compute fbank feat
-    ./local/feat.sh
-fi
-
-if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
-    # decode with fbank feat input
-    ./local/decode.sh
-fi
-
-if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     # decode with wav input
     ./local/recognizer.sh
 fi
 
-if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     # decode with wav input with quanted model
     ./local/recognizer_quant.sh
 fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    # decode with wfst 
+    ./local/recognizer_wfst.sh
+fi