def g2p_en():
clf = RNNG2PModel(EN_G2P_DICT_PATH,
40,
language="en",
rnn=LSTM,
units1=256,
dropout=0.5)
clf.build()
clf.train(G2P_CURRENT_MODEL_DIR, enable_checkpoints=True)
def get_g2p_model(self, language="ru", model_path=None):
if self.g2p_models.get(language) is None:
self.g2p_models[language] = RNNG2PModel(language=language)
if language == "ru" and model_path is None:
model_path = RU_G2P_DEFAULT_MODEL
elif language == "en" and model_path is None:
model_path = EN_G2P_DEFAULT_MODEL
else:
return None
self.g2p_models[language].load(model_path)
return self.g2p_models[language]
def g2p_ru():
clf = RNNG2PModel(RU_G2P_DICT_PATH,
30,
language="ru",
rnn=LSTM,
units1=512,
units2=512,
dropout=0.4,
batch_size=128,
emb_dimension=50)
clf.build()
clf.train(G2P_CURRENT_MODEL_DIR, enable_checkpoints=True)
def __init__(self, language: str = "ru", g2p_model_path: str = None):
self.language = language
self.g2p_model_path = g2p_model_path
if language == "ru":
self.__init_language_defaults(RU_G2P_DEFAULT_MODEL)
elif language == "en":
self.__init_language_defaults(EN_G2P_DEFAULT_MODEL)
else:
raise RuntimeError("Wrong language")
if not os.path.exists(self.g2p_model_path):
raise RuntimeError("No g2p model available (or wrong path)")
self.g2p_model = RNNG2PModel(language=language)
self.g2p_model.load(self.g2p_model_path)
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 convert_to_g2p_only(dict_file, g2p_dict_path, g2p_model):
from rupo.g2p.rnn import RNNG2PModel
g2p_predictor = RNNG2PModel()
g2p_predictor.load(g2p_model)
with open(dict_file, 'r', encoding='utf-8') as r:
lines = r.readlines()
with open(g2p_dict_path, 'w', encoding='utf-8') as w:
words = []
for line in lines:
for word in line.split("#")[1].split(","):
word = word.strip()
clean_word = ""
for i, ch in enumerate(word):
if ch == "'" or ch == "`":
continue
clean_word += ch
words.append(clean_word)
phonetic_words = g2p_predictor.predict(words)
for i, word in enumerate(words):
w.write(word + "\t" + phonetic_words[i] + "\n")
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)