Remove preemphasis from audio processing

datasets/LJSpeech.py

@@ -16,7 +16,7 @@ class LJSpeechDataset(Dataset):
     def __init__(self, csv_file, root_dir, outputs_per_step, sample_rate,
                  text_cleaner, num_mels, min_level_db, frame_shift_ms,
                  frame_length_ms, preemphasis, ref_level_db, num_freq, power,
-                 min_seq_len=0):
+                 min_mel_freq, max_mel_freq, min_seq_len=0):
 
         with open(csv_file, "r", encoding="utf8") as f:
             self.frames = [line.split('|') for line in f]
@@ -26,7 +26,8 @@ class LJSpeechDataset(Dataset):
         self.cleaners = text_cleaner
         self.min_seq_len = min_seq_len
         self.ap = AudioProcessor(sample_rate, num_mels, min_level_db, frame_shift_ms,
-                                 frame_length_ms, preemphasis, ref_level_db, num_freq, power)
+                                 frame_length_ms, preemphasis, ref_level_db, num_freq, power,
+                                 min_mel_freq, max_mel_freq)
         print(" > Reading LJSpeech from - {}".format(root_dir))
         print(" | > Number of instances : {}".format(len(self.frames)))
         self._sort_frames()

train.py

@@ -352,6 +352,8 @@ def main(args):
                                     c.ref_level_db,
                                     c.num_freq,
                                     c.power,
+                                    c.min_mel_freq,
+                                    c.max_mel_freq,
                                     min_seq_len=c.min_seq_len
                                     )
 
@@ -372,7 +374,9 @@ def main(args):
                                   c.preemphasis,
                                   c.ref_level_db,
                                   c.num_freq,
-                                  c.power
+                                  c.power,
+                                  c.min_mel_freq,
+                                  c.max_mel_freq
                                   )
 
     val_loader = DataLoader(val_dataset, batch_size=c.eval_batch_size,

utils/audio.py

@@ -59,11 +59,11 @@ class AudioProcessor(object):
     def _db_to_amp(self, x):
         return np.power(10.0, x * 0.05)
 
-    def apply_preemphasis(self, x):
+    # def apply_preemphasis(self, x):
-        return signal.lfilter([1, -self.preemphasis], [1], x)
+    #     return signal.lfilter([1, -self.preemphasis], [1], x)
-
+    #
-    def apply_inv_preemphasis(self, x):
+    # def apply_inv_preemphasis(self, x):
-        return signal.lfilter([1], [1, -self.preemphasis], x)
+    #     return signal.lfilter([1], [1, -self.preemphasis], x)
 
     def spectrogram(self, y):
         # D = self._stft(self.apply_preemphasis(y))
@@ -105,7 +105,7 @@ class AudioProcessor(object):
         return y
 
     def melspectrogram(self, y):
-        D = self._stft(self.apply_preemphasis(y))
+        D = self._stft(y)
         S = self._amp_to_db(self._linear_to_mel(np.abs(D))) - self.ref_level_db
         return self._normalize(S)
 

utils/generic_utils.py

@@ -40,7 +40,7 @@ def create_experiment_folder(root_path, model_name, debug):
     date_str = datetime.datetime.now().strftime("%B-%d-%Y_%I:%M%p")
     if debug:
         commit_hash = 'debug'
-    else: 
+    else:
         commit_hash = get_commit_hash()
     output_folder = os.path.join(root_path, date_str + '-' + model_name + '-' + commit_hash)
     os.makedirs(output_folder, exist_ok=True)
@@ -135,21 +135,6 @@ def lr_decay(init_lr, global_step, warmup_steps):
     return lr
 
 
-def create_attn_mask(N, T, g=0.05):
-    r'''creating attn mask for guided attention
-    TODO: vectorize'''
-    M = np.zeros([N, T])
-    for t in range(T):
-        for n in range(N):
-            val = 20 * np.exp(-pow((n/N)-(t/T), 2.0)/g)
-            M[n, t] = val
-    e_x = np.exp(M - np.max(M))
-    M = e_x / e_x.sum(axis=0) # only difference
-    M = torch.FloatTensor(M).t().cuda()
-    M = torch.stack([M]*32)
-    return M
-
-
 def mk_decay(init_mk, max_epoch, n_epoch):
     return init_mk * ((max_epoch - n_epoch) / max_epoch)
 
@@ -159,6 +144,20 @@ def count_parameters(model):
     return sum(p.numel() for p in model.parameters() if p.requires_grad)
 
 
+# from https://gist.github.com/jihunchoi/f1434a77df9db1bb337417854b398df1
+def sequence_mask(sequence_length, max_len=None):
+    if max_len is None:
+        max_len = sequence_length.data.max()
+    batch_size = sequence_length.size(0)
+    seq_range = torch.arange(0, max_len).long()
+    seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len)
+    if sequence_length.is_cuda:
+        seq_range_expand = seq_range_expand.cuda()
+    seq_length_expand = (sequence_length.unsqueeze(1)
+                         .expand_as(seq_range_expand))
+    return seq_range_expand < seq_length_expand
+
+
 class Progbar(object):
     """Displays a progress bar.
     Args:

utils/model.py

@@ -1,9 +0,0 @@
-
-def get_param_size(model):
-    params = 0
-    for p in model.parameters():
-        tmp = 1
-        for x in p.size():
-            tmp *= x
-        params += tmp
-    return params