import re
# 音素 (+pau/sil)
phonemes = [
"A",
"E",
"I",
"N",
"O",
"U",
"a",
"b",
"by",
"ch",
"cl",
"d",
"dy",
"e",
"f",
"g",
"gy",
"h",
"hy",
"i",
"j",
"k",
"ky",
"m",
"my",
"n",
"ny",
"o",
"p",
"py",
"r",
"ry",
"s",
"sh",
"t",
"ts",
"ty",
"u",
"v",
"w",
"y",
"z",
"pau",
"sil",
]
extra_symbols = [
"^", # 文の先頭を表す特殊記号 <SOS>
"$", # 文の末尾を表す特殊記号 <EOS> (通常)
"?", # 文の末尾を表す特殊記号 <EOS> (疑問系)
"_", # ポーズ
"#", # アクセント句境界
"[", # ピッチの上がり位置
"]", # ピッチの下がり位置
]
_pad = "~"
# NOTE: 0 をパディングを表す数値とする
symbols = [_pad] + extra_symbols + phonemes
_symbol_to_id = {s: i for i, s in enumerate(symbols)}
_id_to_symbol = {i: s for i, s in enumerate(symbols)}
def numeric_feature_by_regex(regex, s):
match = re.search(regex, s)
if match is None:
return -50
return int(match.group(1))
[docs]def pp_symbols(labels, drop_unvoiced_vowels=True):
"""Extract phoneme + prosoody symbol sequence from input full-context labels
The algorithm is based on [Kurihara 2021] [1]_ with some tweaks.
Args:
labels (HTSLabelFile): List of labels
drop_unvoiced_vowels (bool): Drop unvoiced vowels. Defaults to True.
Returns:
list: List of phoneme + prosody symbols
.. ipython::
In [11]: import ttslearn
In [12]: from nnmnkwii.io import hts
In [13]: from ttslearn.tacotron.frontend.openjtalk import pp_symbols
In [14]: labels = hts.load(ttslearn.util.example_label_file())
In [15]: " ".join(pp_symbols(labels.contexts))
Out[15]: '^ m i [ z u o # m a [ r e ] e sh i a k a r a ... $'
.. [1] K. Kurihara, N. Seiyama, and T. Kumano, “Prosodic features control by
symbols as input of sequence-to-sequence acoustic modeling for neural tts,”
IEICE Transactions on Information and Systems, vol. E104.D, no. 2,
pp. 302–311, 2021.
"""
PP = []
N = len(labels)
# 各音素毎に順番に処理
for n in range(N):
lab_curr = labels[n]
# 当該音素
p3 = re.search(r"\-(.*?)\+", lab_curr).group(1) # type: ignore
# 無声化母音を通常の母音として扱う
if drop_unvoiced_vowels and p3 in "AEIOU":
p3 = p3.lower()
# 先頭と末尾の sil のみ例外対応
if p3 == "sil":
assert n == 0 or n == N - 1
if n == 0:
PP.append("^")
elif n == N - 1:
# 疑問系かどうか
e3 = numeric_feature_by_regex(r"!(\d+)_", lab_curr)
if e3 == 0:
PP.append("$")
elif e3 == 1:
PP.append("?")
continue
elif p3 == "pau":
PP.append("_")
continue
else:
PP.append(p3)
# アクセント型および位置情報（前方または後方）
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)
# アクセント句におけるモーラ数
f1 = numeric_feature_by_regex(r"/F:(\d+)_", lab_curr)
a2_next = numeric_feature_by_regex(r"\+(\d+)\+", labels[n + 1])
# アクセント句境界
if a3 == 1 and a2_next == 1 and p3 in "aeiouAEIOUNcl":
PP.append("#")
# ピッチの立ち下がり（アクセント核）
elif a1 == 0 and a2_next == a2 + 1 and a2 != f1:
PP.append("]")
# ピッチの立ち上がり
elif a2 == 1 and a2_next == 2:
PP.append("[")
return PP
[docs]def num_vocab():
"""Get number of vocabraries
Returns:
int: Number of vocabraries
Examples:
>>> from ttslearn.tacotron.frontend.openjtalk import num_vocab
>>> num_vocab()
>>> 52
"""
return len(symbols)
[docs]def text_to_sequence(text):
"""Convert phoneme + prosody symbols to sequence of numbers
Args:
text (list): text as a list of phoneme + prosody symbols
Returns:
list: List of numbers
Examples:
>>> from ttslearn.tacotron.frontend.openjtalk import text_to_sequence
>>> text_to_sequence(["^", "m", "i", "[", "z","o", "$"])
>>> [1, 31, 27, 6, 49, 35, 2]
"""
return [_symbol_to_id[s] for s in text]
[docs]def sequence_to_text(seq):
"""Convert sequence of numbers to phoneme + prosody symbols
Args:
seq (list): Input sequence of numbers
Returns:
list: List of phoneme + prosody symbols
Examples:
>>> from ttslearn.tacotron.frontend.openjtalk import sequence_to_text
>>> sequence_to_text([1, 31, 27, 6, 49, 35, 2])
>>> ['^', 'm', 'i', '[', 'z', 'o', '$']
"""
return [_id_to_symbol[s] for s in seq]