VoiceAPI / src /tokenizer.py
Harshil748's picture
Add TTS Tokenizer, Technical Report, and Basic Tests
89a8916
Raw
History Blame Contribute Delete
6.48 kB
"""
TTS Tokenizer for VITS models
Adapted from Coqui TTS for SYSPIN models
CRITICAL: The vocabulary MUST be built as:
[<PAD>] + list(punctuations) + list(characters) + [<BLNK>]
Where:
- punctuations = "!¡'(),-.:;¿? " (standard VITS punctuations)
- characters = content of chars.txt file
"""
import re
from typing import Dict, List, Optional
from dataclasses import dataclass
# Standard VITS punctuations used by SYSPIN models
VITS_PUNCTUATIONS = "!¡'(),-.:;¿? "
@dataclass
class CharactersConfig:
"""Character configuration for tokenizer"""
characters: str = ""
punctuations: str = VITS_PUNCTUATIONS
pad: str = "<PAD>"
eos: str = None
bos: str = None
blank: str = "<BLNK>"
phonemes: Optional[str] = None
class TTSTokenizer:
"""
Tokenizer for TTS models - Compatible with SYSPIN VITS models.
The vocabulary is built EXACTLY as VitsCharacters._create_vocab():
vocab = [<PAD>] + list(punctuations) + list(characters) + [<BLNK>]
For SYSPIN models:
- punctuations = "!¡'(),-.:;¿? " (13 chars)
- characters = content from chars.txt
- Total vocab = 1 + 13 + len(chars.txt) + 1
"""
def __init__(
self,
characters: str,
punctuations: str = VITS_PUNCTUATIONS,
pad: str = "<PAD>",
blank: str = "<BLNK>",
):
"""
Initialize tokenizer.
Args:
characters: The characters string (from chars.txt)
punctuations: Punctuation characters (default: VITS standard)
pad: Padding token
blank: Blank token for CTC
"""
self.characters = characters
self.punctuations = punctuations
self.pad = pad
self.blank = blank
# Build vocabulary: [PAD] + punctuations + characters + [BLANK]
self._build_vocab()
def _build_vocab(self):
"""
Build vocabulary EXACTLY matching VitsCharacters._create_vocab():
self._vocab = [self._pad] + list(self._punctuations) + list(self._characters) + [self._blank]
"""
self.vocab: List[str] = []
self.char_to_id: Dict[str, int] = {}
self.id_to_char: Dict[int, str] = {}
# Build vocab in exact order
# 1. PAD token
self.vocab.append(self.pad)
# 2. Punctuations
for char in self.punctuations:
self.vocab.append(char)
# 3. Characters from chars.txt
for char in self.characters:
self.vocab.append(char)
# 4. BLANK token
self.vocab.append(self.blank)
# Build mappings
for idx, char in enumerate(self.vocab):
self.char_to_id[char] = idx
self.id_to_char[idx] = char
self.vocab_size = len(self.vocab)
self.pad_id = self.char_to_id[self.pad]
self.blank_id = self.char_to_id[self.blank]
def text_to_ids(self, text: str, add_blank: bool = True) -> List[int]:
"""
Convert text to token IDs with interspersed blanks.
Matches TTSTokenizer.text_to_ids() from extra.py:
1. Clean text with multilingual_cleaners
2. Encode to IDs
3. Intersperse blank tokens
"""
# Apply multilingual_cleaners
text = self._clean_text(text)
# Encode characters to IDs
char_ids = []
for char in text:
if char in self.char_to_id:
char_ids.append(self.char_to_id[char])
# Skip unknown characters (matching original behavior)
# Intersperse blank tokens
if add_blank:
result = [self.blank_id] * (len(char_ids) * 2 + 1)
result[1::2] = char_ids
return result
return char_ids
def ids_to_text(self, ids: List[int]) -> str:
"""Convert token IDs back to text"""
chars = []
for idx in ids:
if idx in self.id_to_char:
char = self.id_to_char[idx]
if char not in [self.pad, self.blank]:
chars.append(char)
return "".join(chars)
def _clean_text(self, text: str) -> str:
"""
Text cleaning matching multilingual_cleaners from extra.py:
1. lowercase
2. replace_symbols
3. remove_aux_symbols
4. collapse_whitespace
"""
text = text.lower()
text = self._replace_symbols(text)
text = self._remove_aux_symbols(text)
text = re.sub(r"\s+", " ", text).strip()
return text
def _replace_symbols(self, text: str) -> str:
"""Replace symbols matching extra.py replace_symbols()"""
text = text.replace(";", ",")
text = text.replace("-", " ")
text = text.replace(":", ",")
return text
def _remove_aux_symbols(self, text: str) -> str:
"""Remove auxiliary symbols matching extra.py remove_aux_symbols()"""
text = re.sub(r"[\<\>\(\)\[\]\"]+", "", text)
return text
@classmethod
def from_chars_file(cls, chars_file: str) -> "TTSTokenizer":
"""
Create tokenizer from chars.txt file.
This matches the jit_infer.py setup:
- characters = content of chars.txt
- punctuations = "!¡'(),-.:;¿? " (standard VITS punctuations)
Vocab will be: [<PAD>] + punctuations + characters + [<BLNK>]
"""
with open(chars_file, "r", encoding="utf-8") as f:
characters = f.read().strip("\n")
return cls(
characters=characters,
punctuations=VITS_PUNCTUATIONS,
pad="<PAD>",
blank="<BLNK>",
)
class TextNormalizer:
"""Text normalizer for Indian languages"""
@staticmethod
def normalize_numbers(text: str, lang: str = "hi") -> str:
"""Convert numbers to words"""
pattern = r"\{(\d+)\}\{([^}]+)\}"
text = re.sub(pattern, r"\2", text)
return text
@staticmethod
def normalize_punctuation(text: str) -> str:
"""Normalize punctuation marks"""
text = re.sub(r'["""]', '"', text)
text = re.sub(r"[''']", "'", text)
text = re.sub(r"[–—]", "-", text)
return text
@staticmethod
def clean_text(text: str, lang: str = "hi") -> str:
"""Full text cleaning pipeline"""
text = TextNormalizer.normalize_numbers(text, lang)
text = TextNormalizer.normalize_punctuation(text)
text = re.sub(r"\s+", " ", text).strip()
return text