Update app.py

app.py

@@ -4,40 +4,42 @@ import edge_tts
 import os
 import re
 import scipy.io.wavfile
+import numpy as np
 from transformers import pipeline
 
 GENAI_KEY = os.getenv("GEMINI_KEY")
 
-# 음악 모델 로드
+# 음악 생성 모델 로드 (Hugging Face 환경 최적화)
 try:
     music_synthesiser = pipeline("text-to-audio", "facebook/musicgen-small")
-except:
+except Exception as e:
+    print(f"Music Model Load Error: {e}")
     music_synthesiser = None
 
-# 8인 멤버 보이스 데이터 (절대 삭제 안 함)
+# 8인 멤버 보이스 데이터 및 프롬프트 (절대 유지)
 MEMBERS = {
-    "서윤 (Korea)": "ko-KR-SunHiNeural",
-    "Chloe (USA)": "en-US-AriaNeural",
-    "Beatrice (Brazil)": "pt-BR-FranciscaNeural",
-    "Naomi (Japan)": "ja-JP-NanamiNeural",
-    "Elena (Spain)": "es-ES-ElviraNeural",
-    "Amira (Egypt)": "ar-EG-SalmaNeural",
-    "Liwei (China)": "zh-CN-XiaoxiaoNeural",
-    "Sophie (France)": "fr-FR-DeniseNeural"
+    "서윤 (Korea)": { "voice": "ko-KR-SunHiNeural", "prompt": "당신은 보컬 '서윤'입니다. 시니컬하지만 음악에 진심입니다." },
+    "Chloe (USA)": { "voice": "en-US-AriaNeural", "prompt": "You are 'Chloe', the lead guitarist. Cool rockstar vibe." },
+    "Beatrice (Brazil)": { "voice": "pt-BR-FranciscaNeural", "prompt": "You are 'Beatrice', the drummer. Focus on rhythm." },
+    "Naomi (Japan)": { "voice": "ja-JP-NanamiNeural", "prompt": "You are 'Naomi'. Focus on harmony." },
+    "Elena (Spain)": { "voice": "es-ES-ElviraNeural", "prompt": "You are 'Elena'. Focus on groove." },
+    "Amira (Egypt)": { "voice": "ar-EG-SalmaNeural", "prompt": "You are 'Amira'. Focus on atmosphere." },
+    "Liwei (China)": { "voice": "zh-CN-XiaoxiaoNeural", "prompt": "You are 'Liwei'. Focus on sound design." },
+    "Sophie (France)": { "voice": "fr-FR-DeniseNeural", "prompt": "You are 'Sophie'. Focus on melody." }
 }
 
 async def band_consulting(user_input, member_name, lang_code):
     try:
-        # ★ 언어 설정 완벽 고정
+        member_data = MEMBERS.get(member_name, MEMBERS["서윤 (Korea)"])
         lang_map = {"ko":"Korean","en":"English","ja":"Japanese","pt":"Portuguese","es":"Spanish","ar":"Arabic","zh":"Chinese","fr":"French"}
         target_lang = lang_map.get(lang_code, "Korean")
         
         system_instruction = f"""
-        당신은 밴드 멤버 '{member_name}'입니다.
-        1. 반드시 '{target_lang}'로만 답변하세요. 다른 언어는 섞지 마세요.
-        2. 음성 답변(TTS)용 본문은 5줄 이내로 핵심만 요약하세요.
-        3. 모든 상세 설명, 조언, 텍스트 악보는 반드시 [TAB] 섹션에 넣으세요.
-        4. 음악 생성용 영어 프롬프트는 마지막에 [MUSIC: 프롬프트] 형식으로 포함하세요.
+        {member_data['prompt']}
+        [LANGUAGE RULE] Respond 100% in {target_lang}. No other languages.
+        [STYLE RULE] 1. Verbal response MUST be under 5 lines (Concise). 
+        2. Put ALL detailed musical explanations, chord progressions, and tips in [TAB].
+        3. End with [MUSIC: English prompt] for a high-quality rock sample.
         """
         
         url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key={GENAI_KEY}"
@@ -54,37 +56,35 @@ async def band_consulting(user_input, member_name, lang_code):
         response = requests.post(url, json=payload)
         res_data = response.json()
         
-        # API 차단 시 Fallback
-        if 'candidates' not in res_data or not res_data['candidates']:
-            clean_text = "이봐 리더, 내 영감이 잠시 필터링됐어. 조금 더 락스타다운 질문으로 다시 해줘!"
-            tab_display = "데이터가 차단되었습니다."
+        if 'candidates' not in res_data:
+            clean_text = "영감이 잠시 차단됐어. 다시 물어봐줘!"
+            tab_display = "필터링으로 인해 내용을 표시할 수 없습니다."
             music_p = "heavy metal rock riff"
         else:
             ai_text_raw = res_data['candidates'][0]['content']['parts'][0]['text']
             tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL)
             music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
-            tab_display = tab_match.group(1).strip() if tab_match else "No Score Available"
+            tab_display = tab_match.group(1).strip() if tab_match else "No Score"
+            
+            # TTS에서 악보 읽기 방지
             clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL)
             clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text).strip()
+            clean_text = "\n".join(clean_text.split('\n')[:5])
             music_p = music_match.group(1).strip() if music_match else "rock guitar riff"
 
-        # ★ 5줄 요약 강제
-        tts_lines = clean_text.split('\n')
-        tts_final = "\n".join(tts_lines[:5])
-
-        # 음성 생성 (악보 기호 완전 제거)
-        tts_input = re.sub(r'[\*\#\-\_\~\|]', '', tts_final)
+        # 음성 생성
+        tts_input = re.sub(r'[\*\#\-\_\~\|]', '', clean_text)
         voice_path = f"/tmp/v_{member_name.split()[0]}.mp3"
-        await edge_tts.Communicate(tts_input, MEMBERS.get(member_name, "ko-KR-SunHiNeural")).save(voice_path)
+        await edge_tts.Communicate(tts_input, member_data['voice']).save(voice_path)
         
-        # 음악 생성 (System Overload 방지)
+        # 음악 생성 (에러 방지 강화)
         music_path = None
         if music_synthesiser:
             try:
                 # 필터링 우회
                 safe_p = music_p.lower()
                 if any(x in safe_p for x in ['laugh', 'vocal', 'voice', 'human']):
-                    safe_p = "distorted electric guitar riff, power rock"
+                    safe_p = "distorted electric guitar, rock style"
                 
                 music_output = music_synthesiser(safe_p, forward_params={"max_new_tokens": 512})
                 if music_output and "audio" in music_output:
@@ -93,12 +93,15 @@ async def band_consulting(user_input, member_name, lang_code):
             except:
                 music_path = None
 
-        return tts_final, voice_path, music_path, tab_display
+        return clean_text, voice_path, music_path, tab_display
+
     except Exception as e:
-        return f"System Error: {str(e)}", None, None, "Error"
+        return f"System Overload: {str(e)}", None, None, "Error"
 
 with gr.Blocks() as demo:
-    i1 = gr.Textbox(); i2 = gr.Textbox(); i3 = gr.Textbox()
+    with gr.Row():
+        i1 = gr.Textbox(); i2 = gr.Textbox(); i3 = gr.Textbox()
     o1 = gr.Textbox(); o2 = gr.Audio(); o3 = gr.Audio(); o4 = gr.Textbox()
     btn = gr.Button("GO"); btn.click(band_consulting, [i1, i2, i3], [o1, o2, o3, o4], api_name="predict")
+
 demo.launch()