File size: 5,975 Bytes
3c7a160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# modified from https://github.com/CjangCjengh/vits/blob/main/text/japanese.py
import re
import sys

import pyopenjtalk


from text import symbols
# Regular expression matching Japanese without punctuation marks:
_japanese_characters = re.compile(
    r"[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]"
)

# Regular expression matching non-Japanese characters or punctuation marks:
_japanese_marks = re.compile(
    r"[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]"
)

# List of (symbol, Japanese) pairs for marks:
_symbols_to_japanese = [(re.compile("%s" % x[0]), x[1]) for x in [("%", "パーセント")]]


# List of (consonant, sokuon) pairs:
_real_sokuon = [
    (re.compile("%s" % x[0]), x[1])
    for x in [
        (r"Q([↑↓]*[kg])", r"k#\1"),
        (r"Q([↑↓]*[tdjʧ])", r"t#\1"),
        (r"Q([↑↓]*[sʃ])", r"s\1"),
        (r"Q([↑↓]*[pb])", r"p#\1"),
    ]
]

# List of (consonant, hatsuon) pairs:
_real_hatsuon = [
    (re.compile("%s" % x[0]), x[1])
    for x in [
        (r"N([↑↓]*[pbm])", r"m\1"),
        (r"N([↑↓]*[ʧʥj])", r"n^\1"),
        (r"N([↑↓]*[tdn])", r"n\1"),
        (r"N([↑↓]*[kg])", r"ŋ\1"),
    ]
]


def post_replace_ph(ph):
    rep_map = {
        ":": ",",
        ";": ",",
        ",": ",",
        "。": ".",
        "!": "!",
        "?": "?",
        "\n": ".",
        "·": ",",
        "、": ",",
        "...": "…",
    }
    if ph in rep_map.keys():
        ph = rep_map[ph]
    if ph in symbols:
        return ph
    if ph not in symbols:
        ph = "UNK"
    return ph


def symbols_to_japanese(text):
    for regex, replacement in _symbols_to_japanese:
        text = re.sub(regex, replacement, text)
    return text


def preprocess_jap(text, with_prosody=False):
    """Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html"""
    text = symbols_to_japanese(text)
    sentences = re.split(_japanese_marks, text)
    marks = re.findall(_japanese_marks, text)
    text = []
    for i, sentence in enumerate(sentences):
        if re.match(_japanese_characters, sentence):
            if with_prosody:
                text += pyopenjtalk_g2p_prosody(sentence)[1:-1]
            else:
                p = pyopenjtalk.g2p(sentence)
                text += p.split(" ")

        if i < len(marks):
            if marks[i] == " ":# 防止意外的UNK
                continue
            text += [marks[i].replace(" ", "")]
    return text


def text_normalize(text):
    # todo: jap text normalize
    return text

# Copied from espnet https://github.com/espnet/espnet/blob/master/espnet2/text/phoneme_tokenizer.py
def pyopenjtalk_g2p_prosody(text, drop_unvoiced_vowels=True):
    """Extract phoneme + prosoody symbol sequence from input full-context labels.

    The algorithm is based on `Prosodic features control by symbols as input of
    sequence-to-sequence acoustic modeling for neural TTS`_ with some r9y9's tweaks.

    Args:
        text (str): Input text.
        drop_unvoiced_vowels (bool): whether to drop unvoiced vowels.

    Returns:
        List[str]: List of phoneme + prosody symbols.

    Examples:
        >>> from espnet2.text.phoneme_tokenizer import pyopenjtalk_g2p_prosody
        >>> pyopenjtalk_g2p_prosody("こんにちは。")
        ['^', 'k', 'o', '[', 'N', 'n', 'i', 'ch', 'i', 'w', 'a', '$']

    .. _`Prosodic features control by symbols as input of sequence-to-sequence acoustic
        modeling for neural TTS`: https://doi.org/10.1587/transinf.2020EDP7104

    """
    labels = pyopenjtalk.make_label(pyopenjtalk.run_frontend(text))
    N = len(labels)

    phones = []
    for n in range(N):
        lab_curr = labels[n]

        # current phoneme
        p3 = re.search(r"\-(.*?)\+", lab_curr).group(1)
        # deal unvoiced vowels as normal vowels
        if drop_unvoiced_vowels and p3 in "AEIOU":
            p3 = p3.lower()

        # deal with sil at the beginning and the end of text
        if p3 == "sil":
            assert n == 0 or n == N - 1
            if n == 0:
                phones.append("^")
            elif n == N - 1:
                # check question form or not
                e3 = _numeric_feature_by_regex(r"!(\d+)_", lab_curr)
                if e3 == 0:
                    phones.append("$")
                elif e3 == 1:
                    phones.append("?")
            continue
        elif p3 == "pau":
            phones.append("_")
            continue
        else:
            phones.append(p3)

        # accent type and position info (forward or backward)
        a1 = _numeric_feature_by_regex(r"/A:([0-9\-]+)\+", lab_curr)
        a2 = _numeric_feature_by_regex(r"\+(\d+)\+", lab_curr)
        a3 = _numeric_feature_by_regex(r"\+(\d+)/", lab_curr)

        # number of mora in accent phrase
        f1 = _numeric_feature_by_regex(r"/F:(\d+)_", lab_curr)

        a2_next = _numeric_feature_by_regex(r"\+(\d+)\+", labels[n + 1])
        # accent phrase border
        if a3 == 1 and a2_next == 1 and p3 in "aeiouAEIOUNcl":
            phones.append("#")
        # pitch falling
        elif a1 == 0 and a2_next == a2 + 1 and a2 != f1:
            phones.append("]")
        # pitch rising
        elif a2 == 1 and a2_next == 2:
            phones.append("[")

    return phones

# Copied from espnet https://github.com/espnet/espnet/blob/master/espnet2/text/phoneme_tokenizer.py
def _numeric_feature_by_regex(regex, s):
    match = re.search(regex, s)
    if match is None:
        return -50
    return int(match.group(1))

def g2p(norm_text, with_prosody=False):
    phones = preprocess_jap(norm_text, with_prosody)
    phones = [post_replace_ph(i) for i in phones]
    # todo: implement tones and word2ph
    return phones


if __name__ == "__main__":
    phones = g2p("こんにちは, hello, AKITOです,よろしくお願いしますね!")
    print(phones)