add LLM translated CMUDICT-KATAKANA dictionary (#1660)

GPT_SoVITS/text/japanese.py

@@ -82,8 +82,9 @@ def post_replace_ph(ph):
         ph = rep_map[ph]
     # if ph in symbols:
     #     return ph
-    # if ph not in symbols:
-    #     ph = "UNK"
+    if ph not in symbols:
+        ph = "UNK"
+        # UNK may be useful as a pause token as it was trained in the model
     return ph
 
 
@@ -103,6 +104,8 @@ def symbols_to_japanese(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)
+    # English words to lower case, should have no influence on japanese words.
+    text = text.lower()
     sentences = re.split(_japanese_marks, text)
     marks = re.findall(_japanese_marks, text)
     text = []
@@ -219,5 +222,6 @@ def g2p(norm_text, with_prosody=True):
 
 
 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！")
+    from text.symbols2 import symbols
+    phones = g2p("Hello.こんにちは！今日もNiCe天気ですね！tokyotowerに行きましょう！")
     print(phones)