Feat: add editor server backend

text/__init__.py

@@ -19,14 +19,25 @@ def cleaned_text_to_sequence(cleaned_text, tones, language):
 
 
 def get_bert(
-    norm_text, word2ph, language, device, assist_text=None, assist_text_weight=0.7
+    text,
+    word2ph,
+    language,
+    device,
+    assist_text=None,
+    assist_text_weight=0.7,
+    ignore_unknown=False,
 ):
     from .chinese_bert import get_bert_feature as zh_bert
     from .english_bert_mock import get_bert_feature as en_bert
     from .japanese_bert import get_bert_feature as jp_bert
 
     lang_bert_func_map = {"ZH": zh_bert, "EN": en_bert, "JP": jp_bert}
-    bert = lang_bert_func_map[language](
-        norm_text, word2ph, device, assist_text, assist_text_weight
-    )
+    if language == "JP":
+        bert = lang_bert_func_map[language](
+            text, word2ph, device, assist_text, assist_text_weight, ignore_unknown
+        )
+    else:
+        bert = lang_bert_func_map[language](
+            text, word2ph, device, assist_text, assist_text_weight
+        )
     return bert

text/cleaner.py

@@ -4,11 +4,13 @@ from text import chinese, japanese, english, cleaned_text_to_sequence
 language_module_map = {"ZH": chinese, "JP": japanese, "EN": english}
 
 
-def clean_text(text, language, use_jp_extra=True):
+def clean_text(text, language, use_jp_extra=True, ignore_unknown=False):
     language_module = language_module_map[language]
     norm_text = language_module.text_normalize(text)
     if language == "JP":
-        phones, tones, word2ph = language_module.g2p(norm_text, use_jp_extra)
+        phones, tones, word2ph = language_module.g2p(
+            norm_text, use_jp_extra, ignore_unknown=ignore_unknown
+        )
     else:
         phones, tones, word2ph = language_module.g2p(norm_text)
     return norm_text, phones, tones, word2ph

text/japanese.py

@@ -2,11 +2,13 @@
 # compatible with Julius https://github.com/julius-speech/segmentation-kit
 import re
 import unicodedata
+from pathlib import Path
 
 import pyopenjtalk
 from num2words import num2words
 from transformers import AutoTokenizer
 
+from common.constants import USER_DICT_PATH, USER_DICT_CSV_PATH
 from common.log import logger
 from text import punctuation
 from text.japanese_mora_list import (
@@ -14,6 +16,11 @@ from text.japanese_mora_list import (
     mora_phonemes_to_mora_kata,
 )
 
+if Path(USER_DICT_CSV_PATH).exists():
+    pyopenjtalk.mecab_dict_index(USER_DICT_CSV_PATH, USER_DICT_PATH)
+if Path(USER_DICT_PATH).exists():
+    pyopenjtalk.update_global_jtalk_with_user_dict(USER_DICT_PATH)
+
 # 子音の集合
 COSONANTS = set(
     [
@@ -160,7 +167,7 @@ def japanese_convert_numbers_to_words(text: str) -> str:
 
 
 def g2p(
-    norm_text: str, use_jp_extra: bool = True
+    norm_text: str, use_jp_extra: bool = True, ignore_unknown: bool = False
 ) -> tuple[list[str], list[int], list[int]]:
     """
     他で使われるメインの関数。`text_normalize()`で正規化された`norm_text`を受け取り、
@@ -182,7 +189,7 @@ def g2p(
 
     # sep_text: 単語単位の単語のリスト
     # sep_kata: 単語単位の単語のカタカナ読みのリスト
-    sep_text, sep_kata = text2sep_kata(norm_text)
+    sep_text, sep_kata = text2sep_kata(norm_text, ignore_unknown=ignore_unknown)
 
     # sep_phonemes: 各単語ごとの音素のリストのリスト
     sep_phonemes = handle_long([kata2phoneme_list(i) for i in sep_kata])
@@ -231,8 +238,8 @@ def g2p(
     return phones, tones, word2ph
 
 
-def g2kata_tone(norm_text: str) -> list[tuple[str, int]]:
-    phones, tones, _ = g2p(norm_text, use_jp_extra=True)
+def g2kata_tone(norm_text: str, ignore_unknown: bool = False) -> list[tuple[str, int]]:
+    phones, tones, _ = g2p(norm_text, use_jp_extra=True, ignore_unknown=ignore_unknown)
     return phone_tone2kata_tone(list(zip(phones, tones)))
 
 
@@ -325,7 +332,9 @@ def g2phone_tone_wo_punct(text: str) -> list[tuple[str, int]]:
     return result
 
 
-def text2sep_kata(norm_text: str) -> tuple[list[str], list[str]]:
+def text2sep_kata(
+    norm_text: str, ignore_unknown: bool = False
+) -> tuple[list[str], list[str]]:
     """
     `text_normalize`で正規化済みの`norm_text`を受け取り、それを単語分割し、
     分割された単語リストとその読み（カタカナor記号1文字）のリストのタプルを返す。
@@ -361,6 +370,9 @@ def text2sep_kata(norm_text: str) -> tuple[list[str], list[str]]:
             # wordは正規化されているので、`.`, `,`, `!`, `'`, `-`, `--` のいずれか
             if not set(word).issubset(set(punctuation)):  # 記号繰り返しか判定
                 # ここはpyopenjtalkが読めない文字等のときに起こる
+                if ignore_unknown:
+                    logger.error(f"Ignoring unknown: {word} in:\n{norm_text}")
+                    continue
                 raise ValueError(f"Cannot read: {word} in:\n{norm_text}")
             # yomiは元の記号のままに変更
             yomi = word
@@ -500,6 +512,9 @@ def handle_long(sep_phonemes: list[list[str]]) -> list[list[str]]:
     おそらく長音記号とダッシュを勘違いしていると思われるので、ダッシュに対応する音素`-`に変換する。
     """
     for i in range(len(sep_phonemes)):
+        if len(sep_phonemes[i]) == 0:
+            # 空白文字等でリストが空の場合
+            continue
         if sep_phonemes[i][0] == "ー":
             if i != 0:
                 prev_phoneme = sep_phonemes[i - 1][-1]

text/japanese_bert.py

@@ -4,7 +4,7 @@ import torch
 from transformers import AutoModelForMaskedLM, AutoTokenizer
 
 from config import config
-from text.japanese import text2sep_kata
+from text.japanese import text2sep_kata, text_normalize
 
 LOCAL_PATH = "./bert/deberta-v2-large-japanese-char-wwm"
 
@@ -19,8 +19,10 @@ def get_bert_feature(
     device=config.bert_gen_config.device,
     assist_text=None,
     assist_text_weight=0.7,
+    ignore_unknown=False,
 ):
-    text = "".join(text2sep_kata(text)[0])
+    text = "".join(text2sep_kata(text, ignore_unknown=ignore_unknown)[0])
+    # text = text_normalize(text)
     if assist_text:
         assist_text = "".join(text2sep_kata(assist_text)[0])
     if (