Update app.py

app.py

@@ -32,9 +32,9 @@ class ModelManager:
             cls._music_pipeline = pipeline("text-to-audio", "facebook/musicgen-small", device="cpu")
         return cls._music_pipeline
 
-# 페르소나 및 멤버 맵 (기존 유지)
+# 멤버별 고유 성격/말투 페르소나
 PERSONA_MAP = {
-    "서윤 (Korea)": "서정적이면서도 강렬한 K-Rock 리더. 진지하고 따뜻하게 조언함.",
+    "서윤 (Korea)": "서정적이면서도 강렬한 K-Rock 리더. 진지하고 철학적이며 따뜻하게 조언함.",
     "Chloe (USA)": "자유분방한 캘리포니아 펑크 락스타. 거침없고 쿨하며 에너제틱한 말투.",
     "Naomi (Japan)": "완벽주의 J-Rock 기타리스트. 섬세하고 예의 바르며 음악 이론에 날카로움.",
     "Beatrice (Brazil)": "열정적인 삼바 락커. 긍정적이고 리듬감이 넘치며 인생의 즐거움을 강조함.",
@@ -44,6 +44,7 @@ PERSONA_MAP = {
     "Sophie (France)": "예술적 자존심이 강한 아방가르드 락커. 시적이고 세련된 표현을 즐겨 씀."
 }
 
+# 언어별 멤버/보이스 매핑
 LANG_MEMBER_MAP = {
     "Korean": {"name": "서윤 (Korea)", "voice": "ko-KR-SunHiNeural"},
     "English": {"name": "Chloe (USA)", "voice": "en-US-AriaNeural"},
@@ -64,18 +65,15 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
         m_info = LANG_MEMBER_MAP.get(selected_lang, LANG_MEMBER_MAP["Korean"])
         persona = PERSONA_MAP.get(m_info['name'], "열정적인 락스타")
         
-        # [지시 수정] 본문은 5줄 이내 요약, 나머지는 [TAB] 섹션에 통합
+        # 본문 요약 및 상세 탭 분리 지시 프롬프트
         system_prompt = f"""You are the rock star '{m_info['name']}'. Respond ONLY in {selected_lang}.
         Persona: {persona}.
 
         CRITICAL RULES:
-        1. MAIN ADVICE: Summarize your advice in exactly 3 to 5 sentences. This will be shown in the main chat.
-        2. [TAB] SECTION: This is where you provide all long, detailed explanations.
-           - Explain the theory, chord progression, and tips in detail here.
-           - Must include Drum patterns (e.g., H|x-x-x-x| S|--o---o-|).
-           - Must include Guitar Solo Tab with fret numbers (e.g., e|--12-15-12--|).
-        3. [TRANSLATION]: English translation of your 3-5 sentence advice only.
-        4. [MUSIC]: English prompt for MusicGen: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}
+        1. MAIN ADVICE: Provide exactly 3 to 5 deep, soulful sentences for the main chat.
+        2. [TAB] SECTION: Put ALL long explanations, music theory, Drum patterns (e.g., H|x-x-x-x| S|--o---o-|), and detailed Guitar Solo tabs here. This must be very technical and professional.
+        3. [TRANSLATION]: English translation of your 3-5 sentence MAIN ADVICE only.
+        4. [MUSIC]: English prompt for MusicGen reflecting: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}
         """
         
         ai_text_raw = ""
@@ -87,7 +85,8 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
                     model="llama-3.3-70b-versatile"
                 )
                 ai_text_raw = res.choices[0].message.content
-            except: pass
+            except Exception as e:
+                print(f"Groq Error: {e}")
 
         if not ai_text_raw:
             qwen = ModelManager.get_qwen()
@@ -95,6 +94,7 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
             out = qwen(input_t, max_new_tokens=1024)
             ai_text_raw = out[0]['generated_text'].split("assistant\n")[-1]
 
+        # 파싱 로직
         tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
         music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
         trans_match = re.search(r'\[TRANSLATION\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
@@ -102,17 +102,19 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
         tab_display = tab_match.group(1).strip() if tab_match else "No Detailed Data"
         eng_translation = trans_match.group(1).strip() if trans_match else ""
 
-        # TTS 음성에서 번역문 제외 로직 강화
+        # 음성에서 번역문/탭/음악프롬프트 완벽 제거
         speech_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL | re.IGNORECASE)
         speech_text = re.sub(r'\[MUSIC:.*?\]', '', speech_text, flags=re.IGNORECASE)
         speech_text = re.sub(r'\[TRANSLATION\].*?(\[|$)', '', speech_text, flags=re.DOTALL | re.IGNORECASE).strip()
         
+        # TTS 생성
         tts_text = re.sub(r'[\*\#\-\_\~\|]', '', speech_text)
         communicate = edge_tts.Communicate(tts_text, m_info["voice"])
         await communicate.save(voice_path)
 
+        # MusicGen 음악 생성
         music_gen = ModelManager.get_music()
-        music_p = music_match.group(1).strip() if music_match else "rock"
+        music_p = music_match.group(1).strip() if music_match else "rock music"
         music_output = music_gen(music_p, forward_params={"max_new_tokens": 512})
         audio_data = np.squeeze(music_output["audio"])
         audio_int16 = (audio_data * 32767).astype(np.int16)
@@ -121,11 +123,18 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
         return speech_text, voice_path, music_path, tab_display, eng_translation
 
     except Exception as e:
-        return f"Error: {str(e)}", None, None, "No Data", "Error"
+        print(f"Final Error: {e}")
+        return f"밴드 시스템 에러: {str(e)}", None, None, "No Data", "Error occurred"
 
 with gr.Blocks() as demo:
     inputs = [gr.Textbox(visible=False) for _ in range(6)]
-    outputs = [gr.Textbox(visible=False), gr.Audio(visible=False), gr.Audio(visible=False), gr.Textbox(visible=False), gr.Textbox(visible=False)]
+    outputs = [
+        gr.Textbox(visible=False), 
+        gr.Audio(visible=False), 
+        gr.Audio(visible=False), 
+        gr.Textbox(visible=False), 
+        gr.Textbox(visible=False)
+    ]
     btn = gr.Button("API", visible=False)
     btn.click(band_consulting, inputs, outputs, api_name="predict")