def test_aligner(self):
aligner = Aligner()
with open(RU_G2P_DICT_PATH, 'r', encoding='utf-8') as r:
lines = r.readlines()[:50]
pairs = [tuple(line.strip().split("\t")) for line in lines]
for g, p in pairs:
print(aligner.align(g, p))
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 __init__(self,
language: str = "ru",
stress_model_path: str = None,
g2p_model_path: str = None,
grapheme_set=RU_GRAPHEME_SET,
g2p_dict_path=None,
aligner_dump_path=None,
ru_wiki_dict=RU_WIKI_DICT,
cmu_dict=CMU_DICT):
self.language = language
self.stress_model_path = stress_model_path
self.g2p_model_path = g2p_model_path
if language == "ru":
self.__init_language_defaults(RU_STRESS_DEFAULT_MODEL,
RU_G2P_DEFAULT_MODEL)
elif language == "en":
self.__init_language_defaults(EN_STRESS_DEFAULT_MODEL,
EN_G2P_DEFAULT_MODEL)
else:
raise RuntimeError("Wrong language")
if not os.path.exists(self.stress_model_path) or not os.path.exists(
self.g2p_model_path):
raise RuntimeError(
"No stress or g2p models available (or wrong paths)")
self.stress_model = RNNStressModel(language=language)
self.stress_model.load(self.stress_model_path)
self.g2p_model = RNNG2PModel(language=language)
self.g2p_model.load(self.g2p_model_path)
self.aligner = Aligner(language,
grapheme_set,
g2p_dict_path,
aligner_dump_path,
ru_wiki_dict=ru_wiki_dict,
cmu_dict=cmu_dict)
def test_aligner(self):
aligner = Aligner()
self.assertEqual(aligner.align('абазия', 'ɐbɐzʲijə'), ('абаз и я', 'ɐbɐzʲijə'))
self.assertEqual(aligner.align('аахенец', 'aəxʲɪnʲɪʦ'), ('аах ен ец', 'aəxʲɪnʲɪʦ'))
self.assertEqual(aligner.align('абатский', 'ɐbaʦkʲɪj'), ('абатск ий', 'ɐbaʦ kʲɪj'))
self.assertEqual(aligner.align('абазинско-русский', 'ɐbɐzʲinskəruskʲɪj'),
('абаз инско-русск ий', 'ɐbɐzʲinskə rus kʲɪj'))
with open(RU_G2P_DICT_PATH, 'r', encoding='utf-8') as r:
lines = r.readlines()[:50]
pairs = [tuple(line.strip().split("\t")) for line in lines]
for g, p in pairs:
logging.debug(aligner.align(g, p))
class RNNStressPredictor(StressPredictor):
def __init__(self,
language: str = "ru",
stress_model_path: str = None,
g2p_model_path: str = None,
grapheme_set=RU_GRAPHEME_SET,
g2p_dict_path=None,
aligner_dump_path=None,
ru_wiki_dict=RU_WIKI_DICT,
cmu_dict=CMU_DICT):
self.language = language
self.stress_model_path = stress_model_path
self.g2p_model_path = g2p_model_path
if language == "ru":
self.__init_language_defaults(RU_STRESS_DEFAULT_MODEL,
RU_G2P_DEFAULT_MODEL)
elif language == "en":
self.__init_language_defaults(EN_STRESS_DEFAULT_MODEL,
EN_G2P_DEFAULT_MODEL)
else:
raise RuntimeError("Wrong language")
if not os.path.exists(self.stress_model_path) or not os.path.exists(
self.g2p_model_path):
raise RuntimeError(
"No stress or g2p models available (or wrong paths)")
self.stress_model = RNNStressModel(language=language)
self.stress_model.load(self.stress_model_path)
self.g2p_model = RNNG2PModel(language=language)
self.g2p_model.load(self.g2p_model_path)
self.aligner = Aligner(language,
grapheme_set,
g2p_dict_path,
aligner_dump_path,
ru_wiki_dict=ru_wiki_dict,
cmu_dict=cmu_dict)
def __init_language_defaults(self, stress_model_path, g2p_model_path):
if self.stress_model_path is None:
self.stress_model_path = stress_model_path
if self.g2p_model_path is None:
self.g2p_model_path = g2p_model_path
def predict(self, word: str) -> List[int]:
word = word.lower()
phonemes = self.g2p_model.predict([word])[0].replace(" ", "")
stresses = self.stress_model.predict([phonemes])[0]
stresses = [
i for i, stress in enumerate(stresses)
if stress == 1 or stress == 2
]
g, p = self.aligner.align(word, phonemes)
stresses = self.aligner.align_stresses(g,
p,
stresses,
is_grapheme=False)
for i, stress in enumerate(stresses):
stresses[i] -= len([ch for ch in g[:stress] if ch == " "])
stresses = [i for i in stresses if i < len(word)]
return stresses