한국어 우선 MeloTTS 웹 서비스 추가 (GPU, 속도/피치, 로그인 없음)

- FastAPI 백엔드: /api/voices, /api/tts, /api/health
- 한국어 메인 + 영어(5종 악센트), 중국어/일본어 UI 제거
- 언어별 목소리 모델 선택, 속도(0.5~2.0x)/피치(±12반음) 조절
- 로그인/사용제한/과금 유도 없음 (무제한 테스트)
- NVIDIA GPU 자동 사용 (torch cu128, Blackwell sm_120)
- 유저 친화 한국어 UI (frontend)
- Docker/compose (GPU 예약), 8788 포트
- MeloTTS(순수 파이썬) 벤더링 + 검증된 버전 고정(constraints)

vendor/melo/text/cleaner.py

@@ -0,0 +1,36 @@
+from . import chinese, japanese, english, chinese_mix, korean, french, spanish
+from . import cleaned_text_to_sequence
+import copy
+
+language_module_map = {"ZH": chinese, "JP": japanese, "EN": english, 'ZH_MIX_EN': chinese_mix, 'KR': korean,
+                    'FR': french, 'SP': spanish, 'ES': spanish}
+
+
+def clean_text(text, language):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    return norm_text, phones, tones, word2ph
+
+
+def clean_text_bert(text, language, device=None):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    
+    word2ph_bak = copy.deepcopy(word2ph)
+    for i in range(len(word2ph)):
+        word2ph[i] = word2ph[i] * 2
+    word2ph[0] += 1
+    bert = language_module.get_bert_feature(norm_text, word2ph, device=device)
+    
+    return norm_text, phones, tones, word2ph_bak, bert
+
+
+def text_to_sequence(text, language):
+    norm_text, phones, tones, word2ph = clean_text(text, language)
+    return cleaned_text_to_sequence(phones, tones, language)
+
+
+if __name__ == "__main__":
+    pass