sound normalization while reading, adapting get_Speaker for multiple datasets

train.py

@@ -43,10 +43,6 @@ print(" > Number of GPUs: ", num_gpus)
 
 
 def setup_loader(ap, is_val=False, verbose=False):
-    global meta_data_train
-    global meta_data_eval
-    if "meta_data_train" not in globals():
-        meta_data_train, meta_data_eval = load_meta_data(c.datasets)
     if is_val and not c.run_eval:
         loader = None
     else:
@@ -470,6 +466,7 @@ def evaluate(model, criterion, criterion_st, ap, global_step, epoch):
 
 # FIXME: move args definition/parsing inside of main?
 def main(args):  # pylint: disable=redefined-outer-name
+    global meta_data_train, meta_data_eval
     # Audio processor
     ap = AudioProcessor(**c.audio)
 
@@ -479,8 +476,12 @@ def main(args):  # pylint: disable=redefined-outer-name
                          c.distributed["backend"], c.distributed["url"])
     num_chars = len(phonemes) if c.use_phonemes else len(symbols)
 
+    # load data instances
+    meta_data_train, meta_data_eval = load_meta_data(c.datasets)
+
+    # parse speakers
     if c.use_speaker_embedding:
-        speakers = get_speakers(c.data_path, c.meta_file_train, c.dataset)
+        speakers = get_speakers(meta_data_train)
         if args.restore_path:
             prev_out_path = os.path.dirname(args.restore_path)
             speaker_mapping = load_speaker_mapping(prev_out_path)

utils/audio.py

@@ -24,6 +24,7 @@ class AudioProcessor(object):
                  clip_norm=True,
                  griffin_lim_iters=None,
                  do_trim_silence=False,
+                 sound_norm=False,
                  **_):
 
         print(" > Setting up Audio Processor...")
@@ -45,6 +46,7 @@ class AudioProcessor(object):
         self.max_norm = 1.0 if max_norm is None else float(max_norm)
         self.clip_norm = clip_norm
         self.do_trim_silence = do_trim_silence
+        self.sound_norm = sound_norm
         self.n_fft, self.hop_length, self.win_length = self._stft_parameters()
         members = vars(self)
         for key, value in members.items():
@@ -243,6 +245,8 @@ class AudioProcessor(object):
             except ValueError:
                 print(f' [!] File cannot be trimmed for silence - {filename}')
         assert self.sample_rate == sr, "%s vs %s"%(self.sample_rate, sr)
+        if self.sound_norm:
+            x = x / x.max() * 0.9
         return x
 
     @staticmethod

utils/speakers.py

@@ -25,9 +25,7 @@ def save_speaker_mapping(out_path, speaker_mapping):
         json.dump(speaker_mapping, f, indent=4)
 
 
-def get_speakers(data_root, meta_file, dataset_type):
+def get_speakers(items):
     """Returns a sorted, unique list of speakers in a given dataset."""
-    preprocessor = get_preprocessor_by_name(dataset_type)
-    items = preprocessor(data_root, meta_file)
     speakers = {e[2] for e in items}
     return sorted(speakers)