fix tone sand_hi bugs for Chinese frontend

pull/2265/head
TianYuan 2 years ago
parent 9292ceb702
commit e1a495a8e6

@ -7,18 +7,18 @@ We use `WER` as an evaluation criterion.
# Start
Run the command below to get the results of the test.
```bash
./run.sh
```
The `avg WER` of g2p is: 0.028952373312476395
The `avg WER` of g2p is: 0.024219452438490413
```text
,--------------------------------------------------------------------.
| ./exp/g2p/text.g2p |
|--------------------------------------------------------------------|
| SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err |
|--------+-----------------+-----------------------------------------|
| Sum/Avg| 9996 299181 | 97.2 2.8 0.0 0.1 2.9 53.3 |
| Sum/Avg| 9996 299181 | 97.6 2.4 0.0 0.0 2.4 49.2 |
`--------------------------------------------------------------------'
```

@ -1359,9 +1359,9 @@ g2pw_onnx_models = {
'G2PWModel': {
'1.0': {
'url':
'https://paddlespeech.bj.bcebos.com/Parakeet/released_models/g2p/G2PWModel.tar',
'https://paddlespeech.bj.bcebos.com/Parakeet/released_models/g2p/G2PWModel_1.0.zip',
'md5':
'63bc0894af15a5a591e58b2130a2bcac',
'7e049a55547da840502cf99e8a64f20e',
},
},
}

@ -31,8 +31,11 @@ from paddlespeech.t2s.frontend.g2pw.dataset import get_char_phoneme_labels
from paddlespeech.t2s.frontend.g2pw.dataset import get_phoneme_labels
from paddlespeech.t2s.frontend.g2pw.dataset import prepare_onnx_input
from paddlespeech.t2s.frontend.g2pw.utils import load_config
from paddlespeech.t2s.frontend.zh_normalization.char_convert import tranditional_to_simplified
from paddlespeech.utils.env import MODEL_HOME
model_version = '1.0'
def predict(session, onnx_input, labels):
all_preds = []
@ -62,34 +65,38 @@ class G2PWOnnxConverter:
style='bopomofo',
model_source=None,
enable_non_tradional_chinese=False):
if not os.path.exists(os.path.join(model_dir, 'G2PWModel/g2pW.onnx')):
uncompress_path = download_and_decompress(
g2pw_onnx_models['G2PWModel']['1.0'], model_dir)
uncompress_path = download_and_decompress(
g2pw_onnx_models['G2PWModel'][model_version], model_dir)
sess_options = onnxruntime.SessionOptions()
sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
sess_options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
sess_options.intra_op_num_threads = 2
self.session_g2pW = onnxruntime.InferenceSession(
os.path.join(model_dir, 'G2PWModel/g2pW.onnx'),
os.path.join(uncompress_path, 'g2pW.onnx'),
sess_options=sess_options)
self.config = load_config(
os.path.join(model_dir, 'G2PWModel/config.py'), use_default=True)
os.path.join(uncompress_path, 'config.py'), use_default=True)
self.model_source = model_source if model_source else self.config.model_source
self.enable_opencc = enable_non_tradional_chinese
self.tokenizer = BertTokenizer.from_pretrained(self.config.model_source)
polyphonic_chars_path = os.path.join(model_dir,
'G2PWModel/POLYPHONIC_CHARS.txt')
monophonic_chars_path = os.path.join(model_dir,
'G2PWModel/MONOPHONIC_CHARS.txt')
polyphonic_chars_path = os.path.join(uncompress_path,
'POLYPHONIC_CHARS.txt')
monophonic_chars_path = os.path.join(uncompress_path,
'MONOPHONIC_CHARS.txt')
self.polyphonic_chars = [
line.split('\t')
for line in open(polyphonic_chars_path, encoding='utf-8').read()
.strip().split('\n')
]
self.non_polyphonic = {
'', '', '', '', '', '', '', '', '', '', '', '', '',
'', '', '', ''
}
self.non_monophonic = {'', ''}
self.monophonic_chars = [
line.split('\t')
for line in open(monophonic_chars_path, encoding='utf-8').read()
@ -101,13 +108,27 @@ class G2PWOnnxConverter:
self.polyphonic_chars)
self.chars = sorted(list(self.char2phonemes.keys()))
self.polyphonic_chars_new = set(self.chars)
for char in self.non_polyphonic:
if char in self.polyphonic_chars_new:
self.polyphonic_chars_new.remove(char)
self.monophonic_chars_dict = {
char: phoneme
for char, phoneme in self.monophonic_chars
}
for char in self.non_monophonic:
if char in self.monophonic_chars_dict:
self.monophonic_chars_dict.pop(char)
self.pos_tags = [
'UNK', 'A', 'C', 'D', 'I', 'N', 'P', 'T', 'V', 'DE', 'SHI'
]
with open(
os.path.join(model_dir,
'G2PWModel/bopomofo_to_pinyin_wo_tune_dict.json'),
os.path.join(uncompress_path,
'bopomofo_to_pinyin_wo_tune_dict.json'),
'r',
encoding='utf-8') as fr:
self.bopomofo_convert_dict = json.load(fr)
@ -117,7 +138,7 @@ class G2PWOnnxConverter:
}[style]
with open(
os.path.join(model_dir, 'G2PWModel/char_bopomofo_dict.json'),
os.path.join(uncompress_path, 'char_bopomofo_dict.json'),
'r',
encoding='utf-8') as fr:
self.char_bopomofo_dict = json.load(fr)
@ -175,25 +196,25 @@ class G2PWOnnxConverter:
return results
def _prepare_data(self, sentences):
polyphonic_chars = set(self.chars)
monophonic_chars_dict = {
char: phoneme
for char, phoneme in self.monophonic_chars
}
texts, query_ids, sent_ids, partial_results = [], [], [], []
for sent_id, sent in enumerate(sentences):
pypinyin_result = pinyin(sent, style=Style.TONE3)
# pypinyin works well for Simplified Chinese than Traditional Chinese
sent_s = tranditional_to_simplified(sent)
pypinyin_result = pinyin(sent_s, style=Style.TONE3)
partial_result = [None] * len(sent)
for i, char in enumerate(sent):
if char in polyphonic_chars:
if char in self.polyphonic_chars_new:
texts.append(sent)
query_ids.append(i)
sent_ids.append(sent_id)
elif char in monophonic_chars_dict:
elif char in self.monophonic_chars_dict:
partial_result[i] = self.style_convert_func(
monophonic_chars_dict[char])
self.monophonic_chars_dict[char])
elif char in self.char_bopomofo_dict:
partial_result[i] = pypinyin_result[i][0]
# partial_result[i] = self.style_convert_func(self.char_bopomofo_dict[char][0])
else:
partial_result[i] = pypinyin_result[i][0]
partial_results.append(partial_result)
return texts, query_ids, sent_ids, partial_results

@ -24,3 +24,21 @@ polyphonic:
撒切尔: ['sa4','qie4','er3']
比比皆是: ['bi3','bi3','jie1','shi4']
身无长物: ['shen1','wu2','chang2','wu4']
手里: ['shou2','li3']
关卡: ['guan1','qia3']
怀揣: ['huai2','chuai1']
挑剔: ['tiao1','ti4']
供称: ['gong4','cheng1']
作坊: ['zuo1', 'fang5']
中医: ['zhong1','yi1']
嚷嚷: ['rang1','rang5']
商厦: ['shang1','sha4']
大厦: ['da4','sha4']
刹车: ['sha1','che1']
嘚瑟: ['de4','se5']
朝鲜: ['chao2','xian3']
阿房宫: ['e1','pang2','gong1']
阿胶: ['e1','jiao1']
咖喱: ['ga1','li5']
时分: ['shi2','fen1']
蚌埠: ['beng4','bu4']

@ -41,30 +41,32 @@ class ToneSandhi():
'棺材', '棒槌', '棉花', '核桃', '栅栏', '柴火', '架势', '枕头', '枇杷', '机灵', '本事',
'木头', '木匠', '朋友', '月饼', '月亮', '暖和', '明白', '时候', '新鲜', '故事', '收拾',
'收成', '提防', '挖苦', '挑剔', '指甲', '指头', '拾掇', '拳头', '拨弄', '招牌', '招呼',
'抬举', '护士', '折腾', '扫帚', '打量', '打算', '', '', '打听', '打发', '扎实',
'扁担', '戒指', '懒得', '意识', '意思', '情形', '悟性', '怪物', '思量', '怎么', '念头',
'念叨', '快活', '忙活', '志气', '心思', '得罪', '张罗', '弟兄', '开通', '应酬', '庄稼',
'干事', '帮手', '帐篷', '希罕', '师父', '师傅', '巴结', '巴掌', '差事', '工夫', '岁数',
'屁股', '尾巴', '少爷', '小气', '小伙', '将就', '对头', '对付', '寡妇', '家伙', '客气',
'实在', '官司', '学问', '学生', '字号', '嫁妆', '媳妇', '媒人', '婆家', '娘家', '委屈',
'姑娘', '姐夫', '妯娌', '妥当', '妖精', '奴才', '女婿', '头发', '太阳', '大爷', '大方',
'大意', '大夫', '多少', '多么', '外甥', '壮实', '地道', '地方', '在乎', '困难', '嘴巴',
'嘱咐', '嘟囔', '嘀咕', '喜欢', '喇嘛', '喇叭', '商量', '唾沫', '哑巴', '哈欠', '哆嗦',
'咳嗽', '和尚', '告诉', '告示', '含糊', '吓唬', '后头', '名字', '名堂', '合同', '吆喝',
'叫唤', '口袋', '厚道', '厉害', '千斤', '包袱', '包涵', '匀称', '勤快', '动静', '动弹',
'功夫', '力气', '前头', '刺猬', '刺激', '别扭', '利落', '利索', '利害', '分析', '出息',
'凑合', '凉快', '冷战', '冤枉', '冒失', '养活', '关系', '先生', '兄弟', '便宜', '使唤',
'佩服', '作坊', '体面', '位置', '似的', '伙计', '休息', '什么', '人家', '亲戚', '亲家',
'交情', '云彩', '事情', '买卖', '主意', '丫头', '丧气', '两口', '东西', '东家', '世故',
'不由', '不在', '下水', '下巴', '上头', '上司', '丈夫', '丈人', '一辈', '那个', '菩萨',
'父亲', '母亲', '咕噜', '邋遢', '费用', '冤家', '甜头', '介绍', '荒唐', '大人', '泥鳅',
'幸福', '熟悉', '计划', '扑腾', '蜡烛', '姥爷', '照顾', '喉咙', '吉他', '弄堂', '蚂蚱',
'凤凰', '拖沓', '寒碜', '糟蹋', '倒腾', '报复', '逻辑', '盘缠', '喽啰', '牢骚', '咖喱',
'扫把', '惦记'
'抬举', '护士', '折腾', '扫帚', '打量', '打算', '', '打听', '打发', '扎实', '扁担',
'戒指', '懒得', '意识', '意思', '情形', '悟性', '怪物', '思量', '怎么', '念头', '念叨',
'快活', '忙活', '志气', '心思', '得罪', '张罗', '弟兄', '开通', '应酬', '庄稼', '干事',
'帮手', '帐篷', '希罕', '师父', '师傅', '巴结', '巴掌', '差事', '工夫', '岁数', '屁股',
'尾巴', '少爷', '小气', '小伙', '将就', '对头', '对付', '寡妇', '家伙', '客气', '实在',
'官司', '学问', '字号', '嫁妆', '媳妇', '媒人', '婆家', '娘家', '委屈', '姑娘', '姐夫',
'妯娌', '妥当', '妖精', '奴才', '女婿', '头发', '太阳', '大爷', '大方', '大意', '大夫',
'多少', '多么', '外甥', '壮实', '地道', '地方', '在乎', '困难', '嘴巴', '嘱咐', '嘟囔',
'嘀咕', '喜欢', '喇嘛', '喇叭', '商量', '唾沫', '哑巴', '哈欠', '哆嗦', '咳嗽', '和尚',
'告诉', '告示', '含糊', '吓唬', '后头', '名字', '名堂', '合同', '吆喝', '叫唤', '口袋',
'厚道', '厉害', '千斤', '包袱', '包涵', '匀称', '勤快', '动静', '动弹', '功夫', '力气',
'前头', '刺猬', '刺激', '别扭', '利落', '利索', '利害', '分析', '出息', '凑合', '凉快',
'冷战', '冤枉', '冒失', '养活', '关系', '先生', '兄弟', '便宜', '使唤', '佩服', '作坊',
'体面', '位置', '似的', '伙计', '休息', '什么', '人家', '亲戚', '亲家', '交情', '云彩',
'事情', '买卖', '主意', '丫头', '丧气', '两口', '东西', '东家', '世故', '不由', '下水',
'下巴', '上头', '上司', '丈夫', '丈人', '一辈', '那个', '菩萨', '父亲', '母亲', '咕噜',
'邋遢', '费用', '冤家', '甜头', '介绍', '荒唐', '大人', '泥鳅', '幸福', '熟悉', '计划',
'扑腾', '蜡烛', '姥爷', '照顾', '喉咙', '吉他', '弄堂', '蚂蚱', '凤凰', '拖沓', '寒碜',
'糟蹋', '倒腾', '报复', '逻辑', '盘缠', '喽啰', '牢骚', '咖喱', '扫把', '惦记', '戏弄',
'将军', '别人'
}
self.must_not_neural_tone_words = {
"男子", "女子", "分子", "原子", "量子", "莲子", "石子", "瓜子", "电子", "人人", "虎虎",
"幺幺"
'男子', '女子', '分子', '原子', '量子', '莲子', '石子', '瓜子', '电子', '人人', '虎虎',
'幺幺', '干嘛', '学子', '哈哈', '数数', '袅袅', '局地', '以下', '娃哈哈', '花花草草', '留得',
'耕地', '想想', '熙熙', '攘攘', '卵子', '死死', '冉冉', '恳恳', '佼佼', '吵吵', '打打',
'考考', '整整', '莘莘'
}
self.punc = ":,;。?!“”‘’':,;.?!"
@ -75,27 +77,24 @@ class ToneSandhi():
# finals: ['ia1', 'i3']
def _neural_sandhi(self, word: str, pos: str,
finals: List[str]) -> List[str]:
if word in self.must_not_neural_tone_words:
return finals
# reduplication words for n. and v. e.g. 奶奶, 试试, 旺旺
for j, item in enumerate(word):
if j - 1 >= 0 and item == word[j - 1] and pos[0] in {
"n", "v", "a"
} and word not in self.must_not_neural_tone_words:
if j - 1 >= 0 and item == word[j - 1] and pos[0] in {"n", "v", "a"}:
finals[j] = finals[j][:-1] + "5"
ge_idx = word.find("")
if len(word) >= 1 and word[-1] in "吧呢啊呐噻嘛吖嗨呐哦哒额滴哩哟喽啰耶喔诶":
if len(word) >= 1 and word[-1] in "吧呢啊呐噻嘛吖嗨呐哦哒额滴哩哟喽啰耶喔诶":
finals[-1] = finals[-1][:-1] + "5"
elif len(word) >= 1 and word[-1] in "的地得":
finals[-1] = finals[-1][:-1] + "5"
# e.g. 走了, 看着, 去过
elif len(word) == 1 and word in "了着过" and pos in {"ul", "uz", "ug"}:
finals[-1] = finals[-1][:-1] + "5"
elif len(word) > 1 and word[-1] in "们子" and pos in {
"r", "n"
} and word not in self.must_not_neural_tone_words:
elif len(word) > 1 and word[-1] in "们子" and pos in {"r", "n"}:
finals[-1] = finals[-1][:-1] + "5"
# e.g. 桌上, 地下, 家里
elif len(word) > 1 and word[-1] in "上下" and pos in {"s", "l", "f"}:
# e.g. 桌上, 地下
elif len(word) > 1 and word[-1] in "上下" and pos in {"s", "l", "f"}:
finals[-1] = finals[-1][:-1] + "5"
# e.g. 上来, 下去
elif len(word) > 1 and word[-1] in "来去" and word[-2] in "上下进出回过起开":
@ -147,7 +146,7 @@ class ToneSandhi():
for i, char in enumerate(word):
if char == "" and i + 1 < len(word):
# "一" before tone4 should be yi2, e.g. 一段
if finals[i + 1][-1] == "4":
if finals[i + 1][-1] in {'4', '5'}:
finals[i] = finals[i][:-1] + "2"
# "一" before non-tone4 should be yi4, e.g. 一天
else:
@ -239,7 +238,12 @@ class ToneSandhi():
for i, (word, pos) in enumerate(seg):
if i - 1 >= 0 and word == "" and i + 1 < len(seg) and seg[i - 1][
0] == seg[i + 1][0] and seg[i - 1][1] == "v":
new_seg[i - 1][0] = new_seg[i - 1][0] + "" + new_seg[i - 1][0]
if i - 1 < len(new_seg):
new_seg[i -
1][0] = new_seg[i - 1][0] + "" + new_seg[i - 1][0]
else:
new_seg.append([word, pos])
new_seg.append([seg[i + 1][0], pos])
else:
if i - 2 >= 0 and seg[i - 1][0] == "" and seg[i - 2][
0] == word and pos == "v":

@ -84,6 +84,24 @@ class Frontend():
self.tone_modifier = ToneSandhi()
self.text_normalizer = TextNormalizer()
self.punc = ":,;。?!“”‘’':,;.?!"
self.phrases_dict = {
'开户行': [['ka1i'], ['hu4'], ['hang2']],
'发卡行': [['fa4'], ['ka3'], ['hang2']],
'放款行': [['fa4ng'], ['kua3n'], ['hang2']],
'茧行': [['jia3n'], ['hang2']],
'行号': [['hang2'], ['ha4o']],
'各地': [['ge4'], ['di4']],
'借还款': [['jie4'], ['hua2n'], ['kua3n']],
'时间为': [['shi2'], ['jia1n'], ['we2i']],
'为准': [['we2i'], ['zhu3n']],
'色差': [['se4'], ['cha1']],
'': [['dia3']],
'': [['bei5']],
'': [['bu4']],
'': [['zuo5']],
'': [['lei5']],
'掺和': [['chan1'], ['huo5']]
}
# g2p_model can be pypinyin and g2pM and g2pW
self.g2p_model = g2p_model
if self.g2p_model == "g2pM":
@ -91,6 +109,8 @@ class Frontend():
self.pinyin2phone = generate_lexicon(
with_tone=True, with_erhua=False)
elif self.g2p_model == "g2pW":
# use pypinyin as backup for non polyphonic characters in g2pW
self._init_pypinyin()
self.corrector = Polyphonic()
self.g2pM_model = G2pM()
self.g2pW_model = G2PWOnnxConverter(
@ -99,8 +119,10 @@ class Frontend():
with_tone=True, with_erhua=False)
else:
self.__init__pypinyin()
self.must_erhua = {"小院儿", "胡同儿", "范儿", "老汉儿", "撒欢儿", "寻老礼儿", "妥妥儿"}
self._init_pypinyin()
self.must_erhua = {
"小院儿", "胡同儿", "范儿", "老汉儿", "撒欢儿", "寻老礼儿", "妥妥儿", "媳妇儿"
}
self.not_erhua = {
"虐儿", "为儿", "护儿", "瞒儿", "救儿", "替儿", "有儿", "一儿", "我儿", "俺儿", "妻儿",
"拐儿", "聋儿", "乞儿", "患儿", "幼儿", "孤儿", "婴儿", "婴幼儿", "连体儿", "脑瘫儿",
@ -108,6 +130,7 @@ class Frontend():
"孙儿", "侄孙儿", "女儿", "男儿", "红孩儿", "花儿", "虫儿", "马儿", "鸟儿", "猪儿", "猫儿",
"狗儿"
}
self.vocab_phones = {}
self.vocab_tones = {}
if phone_vocab_path:
@ -121,20 +144,9 @@ class Frontend():
for tone, id in tone_id:
self.vocab_tones[tone] = int(id)
def __init__pypinyin(self):
def _init_pypinyin(self):
large_pinyin.load()
load_phrases_dict({u'开户行': [[u'ka1i'], [u'hu4'], [u'hang2']]})
load_phrases_dict({u'发卡行': [[u'fa4'], [u'ka3'], [u'hang2']]})
load_phrases_dict({u'放款行': [[u'fa4ng'], [u'kua3n'], [u'hang2']]})
load_phrases_dict({u'茧行': [[u'jia3n'], [u'hang2']]})
load_phrases_dict({u'行号': [[u'hang2'], [u'ha4o']]})
load_phrases_dict({u'各地': [[u'ge4'], [u'di4']]})
load_phrases_dict({u'借还款': [[u'jie4'], [u'hua2n'], [u'kua3n']]})
load_phrases_dict({u'时间为': [[u'shi2'], [u'jia1n'], [u'we2i']]})
load_phrases_dict({u'为准': [[u'we2i'], [u'zhu3n']]})
load_phrases_dict({u'色差': [[u'se4'], [u'cha1']]})
load_phrases_dict(self.phrases_dict)
# 调整字的拼音顺序
load_single_dict({ord(u''): u'de,di4'})
@ -258,7 +270,6 @@ class Frontend():
phones.append('sp')
if v and v not in self.punc:
phones.append(v)
phones_list.append(phones)
if merge_sentences:
merge_list = sum(phones_list, [])
@ -275,6 +286,10 @@ class Frontend():
finals: List[str],
word: str,
pos: str) -> List[List[str]]:
# fix er1
for i, phn in enumerate(finals):
if i == len(finals) - 1 and word[i] == "" and phn == 'er1':
finals[i] = 'er2'
if word not in self.must_erhua and (word in self.not_erhua or
pos in {"a", "j", "nr"}):
return initials, finals

Loading…
Cancel
Save