Phonemes WIP

tts/phonetics/base.py

@@ -0,0 +1,16 @@
+"""
+Base class for Arpabet Sequence Transformers.
+"""
+
+class ArpabetTransformer:
+
+    def transform_text(self, text):
+        """
+        Input: String text sequence
+        Output: 2-dimensional tensor sequence
+        """
+        pass
+
+
+
+

tts/phonetics/english_arpabet_v1.py

@@ -0,0 +1,29 @@
+import torch
+import numpy as np
+
+from base import ArpabetTransformer
+
+class EnglishMixedArpabetTransformerV1(ArpabetTransformer):
+
+    def transform_text(self, text):
+        pre_sequence = self._handle_multi_lines(text)
+        return self._convert_to_sequence(pre_sequence)
+
+    def _handle_multi_lines(self, text):
+        # NB: This code is adapted from vanilla Tacotron. It's not very graceful.
+        pre_sequence = ''
+        for line in text.split("\n"):
+            if len(line) < 1:
+                continue
+            if line[-1] != ";":
+                line = line + ";"
+            pre_sequence += line
+
+        return pre_sequence
+
+    def _convert_to_sequence(self, pre_sequence_text):
+        with torch.no_grad():
+            # save VRAM by not including gradients
+            sequence = np.array(text_to_sequence(pre_sequence_text, ['english_cleaners']))[None, :]
+            sequence = torch.autograd.Variable(torch.from_numpy(sequence)).cuda().long()
+        return sequence