class DictStressPredictor(StressPredictor):
def __init__(self,
language="ru",
raw_dict_path=None,
trie_path=None,
zalyzniak_dict=ZALYZNYAK_DICT,
cmu_dict=CMU_DICT):
self.stress_dict = StressDict(language,
raw_dict_path=raw_dict_path,
trie_path=trie_path,
zalyzniak_dict=zalyzniak_dict,
cmu_dict=cmu_dict)
def predict(self, word: str) -> List[int]:
"""
Определение ударения в слове по словарю. Возможно несколько вариантов ударения.
:param word: слово для простановки ударений.
:return stresses: позиции букв, на которые падает ударение.
"""
stresses = []
if count_vowels(word) == 0:
# Если гласных нет, то и ударений нет.
pass
elif count_vowels(word) == 1:
# Если одна гласная, то на неё и падает ударение.
stresses.append(get_first_vowel_position(word))
elif word.find("ё") != -1:
# Если есть буква "ё", то только на неё может падать ударение.
stresses.append(word.find("ё"))
else:
# Проверяем словарь на наличие форм с ударениями.
stresses = self.stress_dict.get_stresses(word, Stress.Type.PRIMARY) +\
self.stress_dict.get_stresses(word, Stress.Type.SECONDARY)
if 'е' not in word:
return stresses
# Находим все возможные варинаты преобразований 'е' в 'ё'.
positions = [i for i in range(len(word)) if word[i] == 'е']
beam = [word[:positions[0]]]
for i in range(len(positions)):
new_beam = []
for prefix in beam:
n = positions[i +
1] if i + 1 < len(positions) else len(word)
new_beam.append(prefix + 'ё' + word[positions[i] + 1:n])
new_beam.append(prefix + 'е' + word[positions[i] + 1:n])
beam = new_beam
# И проверяем их по словарю.
for permutation in beam:
if len(self.stress_dict.get_stresses(permutation)) != 0:
yo_pos = permutation.find("ё")
if yo_pos != -1:
stresses.append(yo_pos)
return stresses
def __init__(self,
language="ru",
raw_dict_path=None,
trie_path=None,
zalyzniak_dict=ZALYZNYAK_DICT,
cmu_dict=CMU_DICT):
self.stress_dict = StressDict(language,
raw_dict_path=raw_dict_path,
trie_path=trie_path,
zalyzniak_dict=zalyzniak_dict,
cmu_dict=cmu_dict)
def convert_to_phoneme_stress(source_file, destination_file, g2p_dict_path,
g2p_model):
from rupo.g2p.rnn import RNNG2PModel
from rupo.g2p.aligner import Aligner
from rupo.stress.dict import StressDict
g2p_predictor = RNNG2PModel(g2p_dict_path)
g2p_predictor.load(g2p_model)
aligner = Aligner()
grapheme_stress_dict_path = os.path.join(
os.path.dirname(os.path.abspath(source_file)),
"ru_grapheme_stress.txt")
ZalyzniakDict.convert_to_accent_only(source_file,
grapheme_stress_dict_path)
d = StressDict(raw_dict_path=grapheme_stress_dict_path)
vowels = set(Phonemes.VOWELS)
with open(destination_file, 'w', encoding='utf-8') as w:
samples = 0
for word, accents in d.get_all():
primary_in_dict = [
int(stress[0]) for stress in accents
if stress[1] == StressDict.StressType.PRIMARY
]
secondary_in_dict = [
int(stress[0]) for stress in accents
if stress[1] == StressDict.StressType.SECONDARY
]
phonemes = g2p_predictor.predict([word])[0]
g, p = aligner.align(word, phonemes)
primary = ZalyzniakDict.align_stresses(g, p, primary_in_dict)
secondary = ZalyzniakDict.align_stresses(
g, p, secondary_in_dict)
is_valid = True
for stress in primary + secondary:
if p[stress] not in vowels:
print(g, p, stress, p[stress])
is_valid = False
if is_valid:
w.write(phonemes + "\t" +
",".join([str(i) for i in primary]) + "\t" +
",".join([str(i) for i in secondary]) + "\n")
samples += 1
if samples % 1000 == 0:
print(samples)
def setUpClass(cls):
cls.dict = StressDict(language="ru",
zalyzniak_dict=ZALYZNYAK_DICT,
raw_dict_path=RU_GRAPHEME_STRESS_PATH,
trie_path=RU_GRAPHEME_STRESS_TRIE_PATH)