fix the bug of spec shape

deepspeech/frontend/featurizer/audio_featurizer.py

@@ -175,6 +175,7 @@ class AudioFeaturizer(object):
                                  max_freq=None,
                                  eps=1e-14):
         """Compute the linear spectrogram from FFT energy."""
+        # return T,D
         if max_freq is None:
             max_freq = sample_rate / 2
         if max_freq > sample_rate / 2:
@@ -190,8 +191,12 @@ class AudioFeaturizer(object):
             window_size=window_size,
             stride_size=stride_size,
             sample_rate=sample_rate)
+
         ind = np.where(freqs <= max_freq)[0][-1] + 1
-        return np.log(specgram[:ind, :] + eps)
+        specgram = np.log(specgram[:ind, :] + eps)
+
+        specgram = np.transpose(specgram) #T,D
+        return specgram
 
     def _specgram_real(self, samples, window_size, stride_size, sample_rate):
         """Compute the spectrogram for samples from a real signal."""
@@ -294,6 +299,7 @@ class AudioFeaturizer(object):
             ceplifter=22,
             useEnergy=True,
             winfunc='povey')
+        
         mfcc_feat = np.transpose(mfcc_feat)
         if delta_delta:
             mfcc_feat = self._concat_delta_delta(mfcc_feat)

deepspeech/frontend/normalizer.py

@@ -131,8 +131,8 @@ class FeatureNormalizer(object):
     def _read_mean_std_from_file(self, filepath, eps=1e-20):
         """Load mean and std from file."""
         mean, istd = load_cmvn(filepath, filetype='json')
-        self._mean = np.expand_dims(mean, axis=-1)
-        self._istd = np.expand_dims(istd, axis=-1)
+        self._mean = np.expand_dims(mean, axis=0)
+        self._istd = np.expand_dims(istd, axis=0)
 
     def write_to_file(self, filepath):
         """Write the mean and stddev to the file.

deepspeech/io/collator.py

@@ -326,10 +326,8 @@ class SpeechCollator():
             audio, text = self.process_feature_and_transform(audio, text)
             #utt
             utts.append(utt)
-            # audio
             # print("---debug---")
             # print(audio.shape)
-            audio=audio.T
             audios.append(audio)  # [T, D]
             audio_lens.append(audio.shape[0])
             # text
@@ -358,7 +356,7 @@ class SpeechCollator():
         self.randomize_feature_parameters(min(audio_lens), n_bins)
         for i in range(len(padded_audios)):
             if not self._randomize_each_batch: 
-                self.randomize_feature_parameters(n_bins, audio_lens[i])
+                self.randomize_feature_parameters(audio_lens[i], n_bins)
             padded_audios[i] = self._augmentation_pipeline.apply_feature_transform(padded_audios[i])
 
         return utts, padded_audios, audio_lens, padded_texts, text_lens

examples/aishell/s0/conf/deepspeech2.yaml

@@ -11,7 +11,7 @@ data:
   max_output_input_ratio: .inf
 
 collator:
-  batch_size: 32 #64 # one gpu
+  batch_size: 64 # one gpu
   randomize_each_batch: False
   mean_std_filepath: data/mean_std.json
   unit_type: char