Add japanese cleaner with prosody info.

GPT_SoVITS/text/japanese.py

@@ -4,8 +4,8 @@ import sys
 
 import pyopenjtalk
 
-from text import symbols
 
+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]"
@@ -71,7 +71,7 @@ def symbols_to_japanese(text):
     return text
 
 
-def preprocess_jap(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)
@@ -79,10 +79,15 @@ def preprocess_jap(text):
     text = []
     for i, sentence in enumerate(sentences):
         if re.match(_japanese_characters, sentence):
-            p = pyopenjtalk.g2p(sentence)
+            if with_prosody:
-            text += p.split(" ")
+                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
 
@@ -91,16 +96,96 @@ 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: str, drop_unvoiced_vowels: bool = True) -> list[str]:
+    """Extract phoneme + prosoody symbol sequence from input full-context labels.
 
-def g2p(norm_text):
+    The algorithm is based on `Prosodic features control by symbols as input of
-    phones = preprocess_jap(norm_text)
+    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__":
-    for line in open("../../../Downloads/transcript_utf8.txt").readlines():
+    phones = g2p("こんにちは, hello, AKITOです,よろしくお願いしますね！")
-        text = line.split(":")[1]
+    print(phones)
-        phones = g2p(text)
-        print(phones)

GPT_SoVITS/text/symbols.py

@@ -315,6 +315,10 @@ ja_symbols = [
     "w",
     "y",
     "z",
+    # "[", #上升调型
+    # "]", #下降调型
+    # "$", #结束符
+    # "^", #开始符
 ]
 
 arpa = {