File size: 12,352 Bytes
b5c6ae8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 | """German front end (fork of the English tiny-voca front end): text -> phoneme tokens with a stress feature.
The same code runs during annotation, training-set preparation and inference,
so tokens can never drift between them. Two programs can produce the phonemes:
* ``espeak`` (front-end version 3, the default): eSpeak NG, as described below.
The current checkpoints and annotations were made with it.
* ``own`` (front-end version 4): the separate ``dizione`` package, a dictionary
built from CMUdict, a small model for unknown words and rules. No eSpeak, so
the whole system can be shipped under a permissive licence. Same 76 symbols.
Choose with ``set_frontend("own")`` or ``TINY_VOCA_FRONTEND=own``. Everything
below about tokens holds for both; the rest of this note is the eSpeak path.
* eSpeak NG (``en-us``) is the only pronunciation source. It is an external
rule-based program: no neural G2P and no downloaded weights.
* A token is one eSpeak phoneme exactly as eSpeak segments it (``--sep``), for
example ``t``, ``aɪ``, ``ɑːɹ`` or ``n̩``. This is also the label inventory of
the phoneme recogniser used for offline forced alignment, so one aligned
label corresponds to exactly one token.
* Lexical stress is a parallel feature (0 none, 1 primary, 2 secondary), not a
token, so it never has to be given audio frames.
* Punctuation, word boundaries and the utterance edges are explicit tokens that
may last zero frames. They carry pauses when the recording has them.
Versions 1 and 2 encoded IPA as UTF-8 bytes and piped text to eSpeak without a
final newline; eSpeak NG 1.51 then drops the last character of the input, so
the final phonemes of every clause were wrong. Version 3 terminates the input
and has a regression test for it.
"""
from dataclasses import dataclass
import hashlib
import json
import os
import re
import shutil
import subprocess
import unicodedata
FRONTEND_VERSION = 103 # German fork: eSpeak NG "de", German phone inventory
ESPEAK_VOICE = "de"
SEPARATOR = "_"
FRONTENDS = ("espeak", "own")
FRONTEND_VARIABLE = "TINY_VOCA_FRONTEND"
_frontend = os.environ.get(FRONTEND_VARIABLE, "espeak")
if _frontend not in FRONTENDS:
raise ValueError(f"{FRONTEND_VARIABLE} must be one of {FRONTENDS}, not {_frontend!r}")
PAD, BOS, EOS, SP, UNK = "<pad>", "<bos>", "<eos>", "<sp>", "<unk>"
SPECIALS = (PAD, BOS, EOS, SP, UNK)
PUNCTUATION = (",", ".", "?", "!", ";", ":", '"', "'", "(", ")", "—", "…")
# German inventory: every phone is a label of the forced-alignment recogniser, except
# "ʊɐ", which eSpeak NG prints (and the recogniser was trained on) as "??".
CONSONANTS = ("p", "b", "t", "d", "k", "ɡ", "f", "v", "s", "z", "ʃ", "ʒ", "ç", "x", "h",
"m", "n", "ŋ", "l", "ɾ", "r", "j", "ts", "pf", "tʃ", "dʒ")
VOWELS = ("ə", "ɐ", "ɜ", "ɪ", "i", "iː", "ɛ", "ɛː", "eː", "a", "ɑ", "ɑː", "ɔ", "oː", "ʊ", "uː",
"y", "yː", "œ", "øː", "aɪ", "aʊ", "ɔø", "ʊɐ")
PHONES = CONSONANTS + VOWELS
SYMBOLS = SPECIALS + PUNCTUATION + PHONES
SYMBOL_TO_ID = {symbol: index for index, symbol in enumerate(SYMBOLS)}
PHONE_SET = frozenset(PHONES)
PHONE_IDS = frozenset(SYMBOL_TO_ID[p] for p in PHONES) | {SYMBOL_TO_ID[UNK]}
VOICELESS = frozenset(("p", "t", "k", "f", "s", "ʃ", "ç", "x", "h", "ts", "pf", "tʃ"))
VOWEL_SET = frozenset(VOWELS)
STRESS_MARKS = {"ˈ": 1, "ˌ": 2}
# eSpeak phonemes outside the inventory (English/French loan words, rare allophones).
FALLBACK = {
"??": ("ʊɐ",), "ɹ": ("ɾ",), "ʁ": ("ɾ",), "w": ("v",), "ð": ("d",), "θ": ("s",), "ʔ": (),
"e": ("eː",), "o": ("oː",), "u": ("uː",), "ø": ("øː",), "ɒ": ("ɔ",), "ʌ": ("a",), "æ": ("ɛ",),
"ɜː": ("øː",), "ɔː": ("oː",), "ɪː": ("iː",), "aː": ("ɑː",), "eɪ": ("eː",), "ɛɪ": ("eː",),
"əʊ": ("oː",), "oʊ": ("oː",), "eə": ("ɛː",), "ɪə": ("iː", "ɐ"), "ʊə": ("uː", "ɐ"), "aɪə": ("aɪ", "ɐ"),
"ɔɪ": ("ɔø",), "ɔʏ": ("ɔø",), "əl": ("ə", "l"), "l̩": ("ə", "l"), "n̩": ("ə", "n"), "m̩": ("ə", "m"),
"ɑ̃": ("ɑ", "ŋ"), "ɔ̃": ("ɔ", "ŋ"), "ɛ̃": ("ɛ", "ŋ"), "œ̃": ("œ", "ŋ"), "ʏ": ("y",), "ɚ": ("ɐ",),
"ɲ": ("n", "j"), "ɬ": ("l",), "ɫ": ("l",), "dz": ("d", "z"), "β": ("v",), "ɣ": ("ɡ",), "ʎ": ("l", "j"),
}
_QUOTES = str.maketrans({"‘": "'", "’": "'", "‚": "'", "‛": "'", "`": "'", "´": "'",
"“": '"', "”": '"', "„": '"', "‟": '"', "«": '"', "»": '"',
"–": "—", "―": "—", "[": "(", "{": "(", "]": ")", "}": ")"})
_PUNCTUATION_CHARS = frozenset(PUNCTUATION)
ABBREVIATIONS = {"dr", "prof", "nr", "str", "bzw", "ca", "usw", "vgl", "evtl", "ggf", "inkl", "hr", "fr", "st", "z", "b", "d", "h", "u", "a", "s", "o"}
_LANGUAGE_TAG = re.compile(r"\((?:_?[a-z0-9]_?){2,3}(?:-(?:_?[a-z0-9]_?)+)*\)")
@dataclass(frozen=True)
class Phonemized:
symbols: tuple
ids: tuple
stress: tuple
unknown: tuple # eSpeak phonemes that had no inventory mapping
def __len__(self):
return len(self.ids)
def frontend():
"""Name of the program that produces the phonemes: ``"espeak"`` or ``"own"``."""
return _frontend
def set_frontend(name):
"""Select the phonemizer for this process. Tokens of the two front ends
must not be mixed: annotations and checkpoints record which one made them."""
global _frontend
if name not in FRONTENDS:
raise ValueError(f"front end must be one of {FRONTENDS}, not {name!r}")
_frontend = name
def _dizione():
try:
import dizione
except ImportError as error:
raise RuntimeError('the "own" front end needs the dizione package: pip install dizione') from error
if tuple(dizione.symbols.SYMBOLS) != SYMBOLS:
raise RuntimeError("dizione's symbol inventory differs from this model's")
return dizione
def frontend_identity():
if _frontend == "own":
phonemizer = _dizione()
return {"frontend_version": phonemizer.VERSION, "phonemizer": "dizione", "symbols": list(SYMBOLS)}
return {"frontend_version": FRONTEND_VERSION, "voice": ESPEAK_VOICE, "separator": SEPARATOR,
"symbols": list(SYMBOLS), "fallback": {key: list(value) for key, value in sorted(FALLBACK.items())}}
def frontend_runtime():
"""What an annotation run records about the exact program behind its
tokens: the eSpeak version, or the fingerprint of the phonemizer build."""
if _frontend == "own":
return {"phonemizer": _dizione().identity()["fingerprint"]}
return {"espeak_version": espeak_version()}
def frontend_fingerprint():
return hashlib.sha256(json.dumps(frontend_identity(), sort_keys=True, ensure_ascii=False).encode()).hexdigest()
def normalize_text(text):
text = unicodedata.normalize("NFKC", text).translate(_QUOTES)
text = re.sub(r"\.{3,}", "…", text)
text = re.sub(r"-{2,}|\s-\s", "—", text)
return re.sub(r"\s+", " ", text).strip()
def chunks(text):
"""Yield ``(is_punctuation, piece)``. Decimals, clock times, digit groups,
contractions, title abbreviations and initialisms stay inside the text piece
for eSpeak; every other punctuation mark becomes its own token."""
start = 0
for i, char in enumerate(text):
if char not in _PUNCTUATION_CHARS:
continue
previous = text[i-1] if i else ""
following = text[i+1] if i+1 < len(text) else ""
if char == "'" and previous.isalnum() and following.isalnum():
continue
if char in ".,:" and previous.isdigit() and following.isdigit():
continue
if char == "." and previous.isdigit() and following == " " and i+2 < len(text) and text[i+2].isalnum():
continue
if char == ".":
word = re.search(r"([A-Za-z]+)$", text[:i])
if word and word[1].lower() in ABBREVIATIONS:
continue
if previous.isalpha() and following.isalpha():
continue
if text[start:i].strip():
yield False, text[start:i].strip()
yield True, char
start = i+1
if text[start:].strip():
yield False, text[start:].strip()
def espeak_executable():
executable = shutil.which("espeak-ng")
if executable is None:
raise RuntimeError("Install espeak-ng on the cloud machine before preprocessing/inference")
return executable
def espeak_version():
return subprocess.run([espeak_executable(), "--version"], capture_output=True, text=True, check=True, timeout=10).stdout.strip()
def espeak_words(piece):
"""Phonemize one text piece. Returns words as lists of ``(phone, stress)``.
The input must end with a newline: eSpeak NG 1.51 silently drops the last
character of unterminated standard input.
"""
result = subprocess.run(
[espeak_executable(), "-q", "--ipa", "--sep=" + SEPARATOR, "-v", ESPEAK_VOICE, "--stdin"],
input=piece.rstrip("\n") + "\n", text=True, capture_output=True, check=True, timeout=30)
return parse_espeak(result.stdout)
def parse_espeak(output):
words = []
for word in _LANGUAGE_TAG.sub("", output).split():
units = []
for unit in word.split(SEPARATOR):
stress = 0
for mark, level in STRESS_MARKS.items():
if mark in unit:
stress = stress or level
unit = unit.replace(mark, "")
unit = "".join(c for c in unit if unicodedata.category(c) != "Cf" and c not in "()")
if unit:
units.append((unit, stress))
if units:
words.append(units)
return words
def canonical(phone):
"""Map an eSpeak phoneme to inventory phones; ``None`` when unknown."""
if phone in PHONE_SET:
return (phone,)
if phone in FALLBACK:
return FALLBACK[phone]
stripped = "".join(c for c in unicodedata.normalize("NFD", phone) if not unicodedata.combining(c))
for candidate in (stripped, stripped.replace("ː", ""), stripped + "ː"):
if candidate != phone and candidate in PHONE_SET:
return (candidate,)
if candidate != phone and candidate in FALLBACK:
return FALLBACK[candidate]
return None
def tokenize(pieces):
"""Build the token sequence from ``(is_punctuation, value)`` pieces where a
text piece's value is the parsed eSpeak word list. Pure function (testable
without eSpeak)."""
symbols, stress, unknown = [BOS], [0], []
for punctuation, value in pieces:
if punctuation:
if symbols[-1] != value: # collapse repeats such as "!!"
symbols.append(value)
stress.append(0)
continue
for word in value:
if symbols[-1] in PHONE_SET or symbols[-1] == UNK:
symbols.append(SP)
stress.append(0)
for phone, level in word:
mapped = canonical(phone)
if mapped is None:
unknown.append(phone)
mapped = (UNK,)
# A fallback may expand to several phones: stress goes to its first vowel.
carrier = next((i for i, item in enumerate(mapped) if item in VOWEL_SET), 0)
for index, item in enumerate(mapped):
symbols.append(item)
stress.append(level if index == carrier else 0)
symbols.append(EOS)
stress.append(0)
if not any(s in PHONE_SET or s == UNK for s in symbols):
raise ValueError("the text has nothing to pronounce")
return Phonemized(tuple(symbols), tuple(SYMBOL_TO_ID[s] for s in symbols), tuple(stress), tuple(unknown))
def phonemize(text):
if not isinstance(text, str) or not text.strip():
raise ValueError("text must not be empty")
if _frontend == "own":
return tokenize(_dizione().pieces(text))
pieces = []
from .normalize_de import normalize as normalize_german
for punctuation, piece in chunks(normalize_text(normalize_german(normalize_text(text)))):
pieces.append((True, piece) if punctuation else (False, espeak_words(piece)))
return tokenize(pieces)
def is_phone(token_ids):
"""Boolean list/tensor helper: tokens that must be realised as audio."""
return [int(i) in PHONE_IDS for i in token_ids]
|