refactor

tts/text/pipeline_components/input_parsing/parser.py

@@ -0,0 +1,56 @@
+"""
+Parse raw text sequences into Arpabet and Grapheme sequences so that they can be 
+treated independently.
+
+Inspired by and including code from original Tacotron2 by Keith Ito.
+"""
+
+# pyright: strict
+
+import re
+from typing import List, Union, Any
+
+# Regular expression matching text enclosed in curly braces:
+_curly_re = re.compile(r'(.*?)\{(.+?)\}(.*)')
+
+class ArpabetSequence:
+    def __init__(self, sequence: str):
+        self.sequence = sequence
+    def __eq__(self, other: Any) -> bool:
+        return isinstance(other, ArpabetSequence) and self.sequence == other.sequence
+
+class GraphemeSequence:
+    def __init__(self, sequence: str):
+        self.sequence = sequence
+    def __eq__(self, other: Any) -> bool:
+        return isinstance(other, GraphemeSequence) and self.sequence == other.sequence
+
+ParsedSequence = List[Union[ArpabetSequence, GraphemeSequence]]
+
+def parse_grapheme_and_arpabet_sequence(input_text: str) -> ParsedSequence:
+    """
+    Parse a string into a list of ArpabetSequence and GraphemeSequence so that each 
+    subsequence can be treated independently.
+    """
+    sequence : ParsedSequence = []
+
+    # Check for curly braces and treat their contents as ARPAbet:
+    while len(input_text):
+        input_text = input_text.strip()
+        m = _curly_re.match(input_text)
+        if not m:
+            sequence.append(GraphemeSequence(input_text))
+            break
+
+        before = m.group(1).strip()
+        bracketed = m.group(2).strip()
+        after = m.group(3).strip()
+
+        if before:
+            sequence.append(GraphemeSequence(before))
+        if bracketed:
+            sequence.append(ArpabetSequence(bracketed))
+
+        input_text = after
+
+    return sequence

tts/text/pipeline_components/input_parsing/test_parser.py

@@ -0,0 +1,45 @@
+from tts.text.pipeline_components.input_parsing.parser import ArpabetSequence, GraphemeSequence, parse_grapheme_and_arpabet_sequence
+
+class TestParser:
+
+    def test_only_text(self):
+        parsed = parse_grapheme_and_arpabet_sequence("This is a test")
+        assert len(parsed) == 1
+        assert isinstance(parsed[0], GraphemeSequence)
+
+    def test_only_arpabet(self):
+        parsed = parse_grapheme_and_arpabet_sequence("{T EH1 S T IH0 NG}")
+        assert len(parsed) == 1
+        assert isinstance(parsed[0], ArpabetSequence)
+
+    def test_text_and_arpabet(self):
+        parsed = parse_grapheme_and_arpabet_sequence("This is a {T EH1 S T}")
+        assert len(parsed) == 2
+        assert isinstance(parsed[0], GraphemeSequence)
+        assert isinstance(parsed[1], ArpabetSequence)
+
+    def test_text_and_arpabet_interleved_1(self):
+        parsed = parse_grapheme_and_arpabet_sequence("{DH IH1 S} is a {T EH1 S T}")
+        assert len(parsed) == 3
+        assert isinstance(parsed[0], ArpabetSequence)
+        assert isinstance(parsed[1], GraphemeSequence)
+        assert isinstance(parsed[2], ArpabetSequence)
+
+    def test_text_and_arpabet_interleved_2(self):
+        parsed = parse_grapheme_and_arpabet_sequence("Check out {DH IH1 S} {T EH1 S T} of the {S IH1 S T AH0 M}")
+        assert len(parsed) == 5
+        assert isinstance(parsed[0], GraphemeSequence)
+        assert isinstance(parsed[1], ArpabetSequence)
+        assert isinstance(parsed[2], ArpabetSequence)
+        assert isinstance(parsed[3], GraphemeSequence)
+        assert isinstance(parsed[4], ArpabetSequence)
+
+    def test_text_and_arpabet_interleved_2(self):
+        parsed = parse_grapheme_and_arpabet_sequence("Check out {DH IH1 S} {T EH1 S T} of the {S IH1 S T AH0 M}.")
+        assert len(parsed) == 6
+        assert isinstance(parsed[0], GraphemeSequence)
+        assert isinstance(parsed[1], ArpabetSequence)
+        assert isinstance(parsed[2], ArpabetSequence)
+        assert isinstance(parsed[3], GraphemeSequence)
+        assert isinstance(parsed[4], ArpabetSequence)
+        assert isinstance(parsed[5], GraphemeSequence)

tts/text/pipeline_components/normalization/base_normalization.py

@@ -0,0 +1,4 @@
+class BaseNormalization:
+
+    def normalize(self, raw_input_text: str) -> str:
+        return raw_input_text

tts/text/pipeline_components/normalization/common_cleaners.py

@@ -0,0 +1,32 @@
+"""
+Copied or adapted from Keith Ito's Tacotron2 code.
+https://github.com/keithito/tacotron
+"""
+
+import re
+from unidecode import unidecode
+
+# Regular expression matching whitespace:
+_whitespace_re = re.compile(r'\s+')
+
+def lowercase(text: str) -> str:
+  return text.lower()
+
+def collapse_whitespace(text) -> str:
+  return re.sub(_whitespace_re, ' ', text)
+
+def convert_to_ascii(text: str) -> str:
+  return unidecode(text)
+
+def basic_cleaners(text: str) -> str:
+  '''Basic pipeline that lowercases and collapses whitespace without transliteration.'''
+  text = lowercase(text)
+  text = collapse_whitespace(text)
+  return text
+
+def transliteration_cleaners(text: str) -> str:
+  '''Pipeline for non-English text that transliterates to ASCII.'''
+  text = convert_to_ascii(text)
+  text = lowercase(text)
+  text = collapse_whitespace(text)
+  return text

tts/text/pipeline_components/normalization/english_normalization.py

@@ -0,0 +1,17 @@
+from tts.text.pipeline_components.normalization.base_normalization import BaseNormalization
+from tts.text.pipeline_components.normalization.common_cleaners import collapse_whitespace
+from tts.text.pipeline_components.normalization.common_cleaners import convert_to_ascii
+from tts.text.pipeline_components.normalization.common_cleaners import lowercase
+
+class EnglishNormalization(BaseNormalization):
+
+    def normalize(self, raw_input_text: str) -> str:
+        text = convert_to_ascii(raw_input_text)
+        text = lowercase(text)
+
+        # TODO: Expand abbreviations
+        # TODO: Expand numbers
+
+        text = collapse_whitespace(text)
+
+        return text

tts/text/pipeline_components/normalization/spanish_normalization.py

@@ -0,0 +1,6 @@
+from tts.text.pipeline_components.normalization.base_normalization import BaseNormalization
+
+class SpanishNormalization(BaseNormalization):
+
+    def normalize(self, raw_input_text: str) -> str:
+        return raw_input_text

tts/text/pipeline_components/normalization/test_english_normalization.py

@@ -0,0 +1,13 @@
+from tts.text.pipeline_components.normalization.english_normalization import EnglishNormalization
+
+class TestEnglishNormalization:
+
+    def test_whitespace_collapse(self):
+        normalizer = EnglishNormalization()
+        output = normalizer.normalize(" \t \n   normalized   \n \t ")
+        assert output == " normalized "
+
+    def test_lowercase(self):
+        normalizer = EnglishNormalization()
+        output = normalizer.normalize("This is a test of Normalization.")
+        assert output == "this is a test of normalization."

tts/text/pipeline_components/phonetization/base_phonetization.py

@@ -0,0 +1,15 @@
+
+import re
+from typing import List
+from tts.text.pipeline_components.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)]

tts/text/pipeline_components/phonetization/cmudict.py

@@ -0,0 +1,72 @@
+""" 
+Contents of this file are mostly from https://github.com/keithito/tacotron
+"""
+
+import re
+
+# TODO: Don't use strings for arpabet phoneme representation, and share the memory.
+
+# Relative to TTS codebase root
+# This does not include any CMUDict extensions or alternatives (eg. for other languages).
+DEFAULT_CMUDICT_PATH = 'dictionaries/cmudict-0.7b'
+
+valid_symbols = [
+  'AA', 'AA0', 'AA1', 'AA2', 'AE', 'AE0', 'AE1', 'AE2', 'AH', 'AH0', 'AH1', 'AH2',
+  'AO', 'AO0', 'AO1', 'AO2', 'AW', 'AW0', 'AW1', 'AW2', 'AY', 'AY0', 'AY1', 'AY2',
+  'B', 'CH', 'D', 'DH', 'EH', 'EH0', 'EH1', 'EH2', 'ER', 'ER0', 'ER1', 'ER2', 'EY',
+  'EY0', 'EY1', 'EY2', 'F', 'G', 'HH', 'IH', 'IH0', 'IH1', 'IH2', 'IY', 'IY0', 'IY1',
+  'IY2', 'JH', 'K', 'L', 'M', 'N', 'NG', 'OW', 'OW0', 'OW1', 'OW2', 'OY', 'OY0',
+  'OY1', 'OY2', 'P', 'R', 'S', 'SH', 'T', 'TH', 'UH', 'UH0', 'UH1', 'UH2', 'UW',
+  'UW0', 'UW1', 'UW2', 'V', 'W', 'Y', 'Z', 'ZH'
+]
+
+_valid_symbol_set = set(valid_symbols)
+
+
+class CMUDict:
+  '''Thin wrapper around CMUDict data. http://www.speech.cs.cmu.edu/cgi-bin/cmudict'''
+  def __init__(self, file_or_path, keep_ambiguous=True):
+    if isinstance(file_or_path, str):
+      with open(file_or_path, encoding='latin-1') as f:
+        entries = _parse_cmudict(f)
+    else:
+      entries = _parse_cmudict(file_or_path)
+    if not keep_ambiguous:
+      entries = {word: pron for word, pron in entries.items() if len(pron) == 1}
+    self._entries = entries
+
+
+  def __len__(self):
+    return len(self._entries)
+
+
+  def lookup(self, word):
+    '''Returns list of ARPAbet pronunciations of the given word.'''
+    return self._entries.get(word.upper())
+
+
+
+_alt_re = re.compile(r'\([0-9]+\)')
+
+
+def _parse_cmudict(file):
+  cmudict = {}
+  for line in file:
+    if len(line) and (line[0] >= 'A' and line[0] <= 'Z' or line[0] == "'"):
+      parts = line.split('  ')
+      word = re.sub(_alt_re, '', parts[0])
+      pronunciation = _get_pronunciation(parts[1])
+      if pronunciation:
+        if word in cmudict:
+          cmudict[word].append(pronunciation)
+        else:
+          cmudict[word] = [pronunciation]
+  return cmudict
+
+
+def _get_pronunciation(s):
+  parts = s.strip().split(' ')
+  for part in parts:
+    if part not in _valid_symbol_set:
+      return None
+  return ' '.join(parts)

tts/text/pipeline_components/phonetization/english_phonetization.py

@@ -0,0 +1,30 @@
+from typing import List
+from tts.text.pipeline_components.input_parsing.parser import ArpabetSequence, GraphemeSequence, ParsedSequence
+from tts.text.pipeline_components.phonetization.base_phonetization import BasePhonetization
+from tts.text.pipeline_components.phonetization.cmudict import CMUDict
+
+class EnglishPhonetization(BasePhonetization):
+
+    def __init__(self, phoneme_dictionary: CMUDict):
+        super().__init__(phoneme_dictionary)
+
+    def try_convert_to_phonemes(self, grapheme_sequence: str) -> ParsedSequence:
+        """
+        Attempt to enrich a grapheme sequence with phonemes where word lookup succeeds.
+        Grapheme symbols will be left in place.
+        The output list will need to be space-separated before inference.
+        """
+        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(ArpabetSequence(maybe_phonemes[0]))
+            else:
+                # TODO: Use g2p prediction to predict the phonemes.
+                converted_words.append(GraphemeSequence(word))
+
+        return converted_words

tts/text/pipeline_components/phonetization/test_cmudict.py

@@ -0,0 +1,15 @@
+from tts.text.pipeline_components.phonetization.cmudict import CMUDict
+from tts.text.pipeline_components.phonetization.cmudict import DEFAULT_CMUDICT_PATH
+
+
+class TestCmuDict:
+
+    def test_lookup_success(self):
+        cmudict = CMUDict(DEFAULT_CMUDICT_PATH)
+        arpabet_sequence = cmudict.lookup('word')
+        assert arpabet_sequence == ['W ER1 D'] # TODO: Split the phonemes!
+
+    def test_lookup_failure(self):
+        cmudict = CMUDict(DEFAULT_CMUDICT_PATH)
+        arpabet_sequence = cmudict.lookup('fakeword')
+        assert arpabet_sequence == None

tts/text/pipeline_components/phonetization/test_english_phonetization.py

@@ -0,0 +1,73 @@
+from tts.text.pipeline_components.phonetization.cmudict import DEFAULT_CMUDICT_PATH, CMUDict
+from tts.text.pipeline_components.phonetization.english_phonetization import EnglishPhonetization
+from tts.text.pipeline_components.input_parsing.parser import ArpabetSequence
+from tts.text.pipeline_components.input_parsing.parser import GraphemeSequence
+
+class TestEnglishPhonetization:
+
+    def test_no_input(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.try_convert_to_phonemes("")
+        assert output == []
+
+    def test_all_words_in_dictionary(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.try_convert_to_phonemes("this is a test")
+        assert output == [
+            ArpabetSequence("DH IH1 S"), 
+            ArpabetSequence("IH1 Z"),
+            ArpabetSequence("AH0"), 
+            ArpabetSequence("T EH1 S T"),
+        ]
+
+    def test_most_words_in_dictionary(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.try_convert_to_phonemes("nonword is a test")
+        assert output == [
+            GraphemeSequence("nonword"), 
+            ArpabetSequence("IH1 Z"), 
+            ArpabetSequence("AH0"), 
+            ArpabetSequence("T EH1 S T"),
+        ]
+
+    def test_punctuation(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.try_convert_to_phonemes("this...")
+        assert output == [
+            ArpabetSequence("DH IH1 S"), 
+            GraphemeSequence("..."),
+        ]
+        output = phonetization.try_convert_to_phonemes("this!")
+        assert output == [
+            ArpabetSequence("DH IH1 S"), 
+            GraphemeSequence("!"),
+        ]
+        output = phonetization.try_convert_to_phonemes("this?")
+        assert output == [
+            ArpabetSequence("DH IH1 S"), 
+            GraphemeSequence("?"),
+        ]
+
+    def test_extra_spaces(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.try_convert_to_phonemes("    this    is    a     test   ")
+        assert output == [
+            ArpabetSequence("DH IH1 S"), 
+            ArpabetSequence("IH1 Z"),
+            ArpabetSequence("AH0"), 
+            ArpabetSequence("T EH1 S T"),
+        ]
+
+    def test_multi_non_word_extra_spaces(self):
+        phoneme_dictionary = CMUDict(DEFAULT_CMUDICT_PATH)
+        phonetization = EnglishPhonetization(phoneme_dictionary)
+        output = phonetization.try_convert_to_phonemes("    nonword    anothernonword   ")
+        assert output == [
+            GraphemeSequence("nonword"), 
+            GraphemeSequence("anothernonword"), 
+        ]

tts/text/pipeline_components/sequencing/english_symbols_v1.py

@@ -0,0 +1,50 @@
+"""
+Originally adapted from https://github.com/keithito/tacotron
+
+This defines the set of symbols used in text input to the model.
+
+This is the set of character encodings that FakeYou and UberDuck models have
+been trained on. We freeze it here as "V1", because these may change in future
+notebooks.
+
+Original documentation: 
+    The default is a set of ASCII characters that works well for English or 
+    text that has been run through Unidecode. For other data, you can modify 
+    _characters. See TRAINING_DATA.md for details.
+"""
+
+from tts.text.pipeline_components.phonetization import cmudict
+from typing import List
+
+# NB: DO NOT CHANGE THE ORDER OF SYMBOLS! 
+# Models are trained on and depend upon the order.
+_pad        = '_'
+_punctuation = '!\'(),.:;? ' #NB: Includes space
+_special = '-'
+_letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
+
+# Prepend "@" to ARPAbet symbols to ensure uniqueness (some are the same as uppercase letters):
+_arpabet = ['@' + s for s in cmudict.valid_symbols]
+
+# Export all symbols:
+symbols = [_pad] + list(_special) + list(_punctuation) + list(_letters) + _arpabet
+
+# Mappings from symbol to numeric ID and vice versa:
+# The ordering **CANNOT** change.
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+_id_to_symbol = {i: s for i, s in enumerate(symbols)}
+
+# Spaces can be injected when joining words.
+SPACE_INTEGER_ENCODING = _symbol_to_id[' ']
+
+
+def symbols_to_sequence(symbols: str) -> List[int]:
+    return [_symbol_to_id[s] for s in symbols if _should_keep_symbol(s)]
+
+
+def arpabet_to_sequence(arpabet_string: str) -> List[int]:
+    return symbols_to_sequence(['@' + s for s in arpabet_string.split()])
+
+
+def _should_keep_symbol(s):
+    return s in _symbol_to_id and s is not '_' and s is not '~'

tts/text/pipeline_components/sequencing/test_english_symbols_v1.py

@@ -0,0 +1,29 @@
+from tts.text.pipeline_components.sequencing.english_symbols_v1 import arpabet_to_sequence, symbols_to_sequence
+
+class TestEnglishSymbols:
+
+    def test_no_input_arpabet(self):
+        output = arpabet_to_sequence('')
+        assert output == []
+
+    def test_no_input_symbols(self):
+        output = symbols_to_sequence('')
+        assert output == []
+
+    def test_arpabet_word(self):
+        # NB: Models are trained on these integer encodings. 
+        # If these tests break, the order of symbols may have changed. 
+        output = arpabet_to_sequence('T EH1 S T')
+        assert output == [133, 94, 131, 133]
+
+    def test_grapheme_word(self):
+        # NB: Models are trained on these integer encodings. 
+        # If these tests break, the order of symbols may have changed. 
+        output = symbols_to_sequence('test')
+        assert output == [57, 42, 56, 57]
+
+    def test_grapheme_sentence(self):
+        # NB: Models are trained on these integer encodings. 
+        # If these tests break, the order of symbols may have changed. 
+        output = symbols_to_sequence('this is a test.')
+        assert output == [57, 45, 46, 56, 11, 46, 56, 11, 38, 11, 57, 42, 56, 57, 7]