From 4926faf42c18386f2b3904b3064f2b25e1113e87 Mon Sep 17 00:00:00 2001
From: Brandon Thomas <echelon@gmail.com>
Date: Sat, 2 Jul 2022 11:24:04 -0400
Subject: [PATCH] Work on phonetization

---
 tts/text/phonetization/base_phonetization.py  | 15 ++++++
 .../phonetization/english_phonetization.py    | 24 ++++++++++
 .../test_english_phonetization.py             | 22 +++++++++
 tts/text/pipeline.py                          | 46 +++++++++++++++++++
 4 files changed, 107 insertions(+)
 create mode 100644 tts/text/phonetization/base_phonetization.py
 create mode 100644 tts/text/phonetization/english_phonetization.py
 create mode 100644 tts/text/phonetization/test_english_phonetization.py
 create mode 100644 tts/text/pipeline.py

diff --git a/tts/text/phonetization/base_phonetization.py b/tts/text/phonetization/base_phonetization.py
new file mode 100644
index 0000000..73f4115
--- /dev/null
+++ b/tts/text/phonetization/base_phonetization.py
@@ -0,0 +1,15 @@
+
+import re
+from typing import List
+from tts.text.phonetization.cmudict import CMUDict
+
+# Regex taken from open source tacotron2 extensions
+_arpabet_split_re = re.compile(r'[^\s]*{.*?}[^\s]*|[\w]+|[^\w\s]+')
+
+class BasePhonetization:
+    def __init__(self, phoneme_dictionary: CMUDict):
+        self.phoneme_dictionary = phoneme_dictionary
+
+    def _split_grapheme_sentence_to_words(self, text_sequence: str) -> List[str]:
+        """Split a grapheme word sequence to individual grapheme words"""
+        return [word for word in _arpabet_split_re.findall(text_sequence)]
diff --git a/tts/text/phonetization/english_phonetization.py b/tts/text/phonetization/english_phonetization.py
new file mode 100644
index 0000000..5c48aac
--- /dev/null
+++ b/tts/text/phonetization/english_phonetization.py
@@ -0,0 +1,24 @@
+from typing import List
+from tts.text.phonetization.base_phonetization import BasePhonetization
+from tts.text.phonetization.cmudict import CMUDict
+
+class EnglishPhonetization(BasePhonetization):
+
+    def __init__(self, phoneme_dictionary: CMUDict):
+        super().__init__(phoneme_dictionary)
+
+    def convert_to_phonemes(self, grapheme_sequence: str) -> List[str]:
+        words = self._split_grapheme_sentence_to_words(grapheme_sequence)
+        converted_words = []
+
+        for word in words:
+            maybe_phonemes = self.phoneme_dictionary.lookup(word)
+            if maybe_phonemes:
+                # NB: Dictionary returns a list of possible matches. 
+                # We'll just use the first.
+                converted_words.append(maybe_phonemes[0])
+            else:
+                # TODO: Use g2p prediction to predict the phonemes.
+                converted_words.append(word)
+
+        return converted_words
diff --git a/tts/text/phonetization/test_english_phonetization.py b/tts/text/phonetization/test_english_phonetization.py
new file mode 100644
index 0000000..28662ea
--- /dev/null
+++ b/tts/text/phonetization/test_english_phonetization.py
@@ -0,0 +1,22 @@
+from tts.text.phonetization.cmudict import DEFAULT_CMUDICT_PATH, CMUDict
+from tts.text.phonetization.english_phonetization import EnglishPhonetization
+
+class TestEnglishPhonetization:
+
+    def test_no_input(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.convert_to_phonemes("")
+        assert output == []
+
+    def test_all_words_in_dictionary(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.convert_to_phonemes("this is a test")
+        assert output == ["DH IH1 S", "IH1 Z", "AH0", "T EH1 S T"]
+
+    def test_most_words_in_dictionary(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.convert_to_phonemes("nonword is a test")
+        assert output == ["nonword", "IH1 Z", "AH0", "T EH1 S T"]
diff --git a/tts/text/pipeline.py b/tts/text/pipeline.py
new file mode 100644
index 0000000..de486a6
--- /dev/null
+++ b/tts/text/pipeline.py
@@ -0,0 +1,46 @@
+from enum import Enum, auto
+
+from tts.text.input_parsing.parser import GraphemeSquence, parse_grapheme_and_arpabet_sequence
+from tts.text.normalization.english_normalization import EnglishNormalization
+from tts.text.normalization.spanish_normalization import SpanishNormalization
+
+
+class PipelineType(Enum):
+    LEGACY_FAKEYOU = auto()
+    ENGLISH_FORCED_ARPABET_V1 = auto()
+    SPANISH_FORCED_ARPABET_V1 = auto()
+
+_ENGLISH_NORMALIZER = EnglishNormalization()
+_SPANISH_NORMALIZER = SpanishNormalization()
+
+def _legacy_fakeyou_pipeline(input_text):
+    """
+    Legacy FakeYou.com did not do Grapheme -> Arpabet conversions automatically.
+    """
+    pass
+
+
+def _english_forced_arpabet_v1(input_text):
+    parsed_sequence = parse_grapheme_and_arpabet_sequence(input_text)
+    for i in range(len(parsed_sequence)):
+        subsequence = parsed_sequence[i]
+        if isinstance(subsequence, GraphemeSquence):
+            normalized = _ENGLISH_NORMALIZER.normalize(subsequence.sequence)
+            parsed_sequence[i] = GraphemeSquence(normalized)
+
+    integer_sequence = []
+    
+
+
+def _spanish_forced_arpabet_v1(input_text):
+    parsed_sequence = parse_grapheme_and_arpabet_sequence(input_text)
+
+
+def text_to_sequence(input_text: str, pipeline_type: PipelineType):
+    if pipeline_type == PipelineType.ENGLISH_FORCED_ARPABET_V1:
+        return _english_forced_arpabet_v1(input_text)
+    elif pipeline_type == PipelineType.SPANISH_FORCED_ARPABET_V1:
+        return _spanish_forced_arpabet_v1(input_text)
+    else:
+        return _legacy_fakeyou_pipeline(input_text)
+