Merge pull request #922 from Jackwaterveg/dev

releace the <space> with ' ' in ctcdecoder

deepspeech/decoders/ctcdecoder/swig/ctc_greedy_decoder.cpp

@@ -53,7 +53,8 @@ std::string ctc_greedy_decoder(
     std::string best_path_result;
     for (size_t i = 0; i < idx_vec.size(); ++i) {
         if (idx_vec[i] != blank_id) {
-            best_path_result += vocabulary[idx_vec[i]];
+            std::string ch = vocabulary[idx_vec[i]];
+            best_path_result += (ch == kSPACE) ? tSPACE : ch;
         }
     }
     return best_path_result;

deepspeech/decoders/ctcdecoder/swig/decoder_utils.cpp

@@ -74,7 +74,8 @@ std::vector<std::pair<double, std::string>> get_beam_search_result(
         // convert index to string
         std::string output_str;
         for (size_t j = 0; j < output.size(); j++) {
-            output_str += vocabulary[output[j]];
+            std::string ch = vocabulary[output[j]];
+            output_str += (ch == kSPACE) ? tSPACE : ch;
         }
         std::pair<double, std::string> output_pair(
             -space_prefixes[i]->approx_ctc, output_str);

deepspeech/decoders/ctcdecoder/swig/decoder_utils.h

@@ -21,6 +21,7 @@
 #include "path_trie.h"
 
 const std::string kSPACE = "<space>";
+const std::string tSPACE = " ";
 const float NUM_FLT_INF = std::numeric_limits<float>::max();
 const float NUM_FLT_MIN = std::numeric_limits<float>::min();
 

deepspeech/exps/deepspeech2/bin/test_hub.py

@@ -56,10 +56,6 @@ class DeepSpeech2Tester_hub():
             cutoff_prob=cfg.cutoff_prob,
             cutoff_top_n=cfg.cutoff_top_n,
             num_processes=cfg.num_proc_bsearch)
-        #replace the '<space>' with ' '
-        result_transcripts = [
-            sentence.replace("<space>", " ") for sentence in result_transcripts
-        ]
 
         return result_transcripts
 

deepspeech/exps/deepspeech2/model.py

@@ -341,11 +341,6 @@ class DeepSpeech2Tester(DeepSpeech2Trainer):
             cutoff_prob=cfg.cutoff_prob,
             cutoff_top_n=cfg.cutoff_top_n,
             num_processes=cfg.num_proc_bsearch)
-        #replace the <space> with ' '
-        result_transcripts = [
-            self._text_featurizer.detokenize(sentence)
-            for sentence in result_transcripts
-        ]
 
         self.autolog.times.stamp()
         self.autolog.times.stamp()