Merge pull request #747 from LittleChenCc/develop

refine the code and correct yaml

deepspeech/io/collator_st.py

@@ -563,7 +563,7 @@ class KaldiPrePorocessedCollator(SpeechCollator):
     @property
     def feature_size(self):
         return self._feat_dim
-    
+
     @property
     def stride_ms(self):
         return self._stride_ms

deepspeech/utils/bleu_score.py

@@ -35,6 +35,7 @@ def bleu(hypothesis, reference):
 
     return sacrebleu.corpus_bleu(hypothesis, reference)
 
+
 def char_bleu(hypothesis, reference):
     """Calculate BLEU. BLEU compares reference text and
     hypothesis text in char-level using scarebleu.
@@ -47,7 +48,8 @@ def char_bleu(hypothesis, reference):
     :type hypothesis: list[str]
     :raises ValueError: If the reference number is zero.
     """
-    hypothesis =[' '.join(list(hyp.replace(' ', ''))) for hyp in hypothesis]
-    reference = [[' '.join(list(ref_i.replace(' ', ''))) for ref_i in ref ]for ref in reference ]
+    hypothesis = [' '.join(list(hyp.replace(' ', ''))) for hyp in hypothesis]
+    reference = [[' '.join(list(ref_i.replace(' ', ''))) for ref_i in ref]
+                 for ref in reference]
 
-    return sacrebleu.corpus_bleu(hypothesis, reference)
\ No newline at end of file
+    return sacrebleu.corpus_bleu(hypothesis, reference)

examples/dataset/ted_en_zh/ted_en_zh.py

@@ -44,9 +44,11 @@ def create_manifest(data_dir, manifest_path_prefix):
     print("Creating manifest %s ..." % manifest_path_prefix)
     json_lines = []
 
-    data_types_infos = [('train', 'train-split/train-segment', 'En-Zh/train.en-zh'), 
-                ('dev', 'test-segment/tst2010', 'En-Zh/tst2010.en-zh'), 
-                ('test', 'test-segment/tst2015', 'En-Zh/tst2015.en-zh')]
+    data_types_infos = [
+        ('train', 'train-split/train-segment', 'En-Zh/train.en-zh'),
+        ('dev', 'test-segment/tst2010', 'En-Zh/tst2010.en-zh'),
+        ('test', 'test-segment/tst2015', 'En-Zh/tst2015.en-zh')
+    ]
     for data_info in data_types_infos:
         dtype, audio_relative_dir, text_relative_path = data_info
         del json_lines[:]
@@ -63,7 +65,7 @@ def create_manifest(data_dir, manifest_path_prefix):
                 continue
             audio_id, trancription, translation = line.split('\t')
             utt = audio_id.split('.')[0]
-            
+
             audio_path = os.path.join(audio_dir, audio_id)
             if os.path.exists(audio_path):
                 if os.path.getsize(audio_path) < 30000:

examples/ted_en_zh/conf/transformer_joint_noam.yaml

@@ -3,8 +3,8 @@ data:
   train_manifest: data/manifest.train
   dev_manifest: data/manifest.dev
   test_manifest: data/manifest.test
-  min_input_len: 0.5  # second
-  max_input_len: 3000.0 # second
+  min_input_len: 0.05  # second
+  max_input_len: 30.0 # second
   min_output_len: 0.0 # tokens
   max_output_len: 400.0 # tokens
   min_output_input_ratio: 0.01