Update app.py

app.py

@@ -32,6 +32,18 @@ class ModelManager:
             cls._music_pipeline = pipeline("text-to-audio", "facebook/musicgen-small", device="cpu")
         return cls._music_pipeline
 
+# 멤버별 고유 성격/말투 페르소나
+PERSONA_MAP = {
+    "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.",
+    "Elena (Spain)": "드라마틱한 감성의 소유자. 플라멩코의 정열과 락의 파괴력을 동시에 가짐.",
+    "Amira (Egypt)": "신비로운 오리엔탈 락커. 깊은 지혜와 고전적인 무게감을 담아 조언함.",
+    "Liwei (China)": "전통과 현대의 조화를 중시하는 퓨전 락커. 절제되고 힘 있는 말투.",
+    "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"},
@@ -51,13 +63,21 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
         music_path = f"/tmp/m_{req_id}.wav"
         
         m_info = LANG_MEMBER_MAP.get(selected_lang, LANG_MEMBER_MAP["Korean"])
+        persona = PERSONA_MAP.get(m_info['name'], "열정적인 락스타")
         
-        # 프롬프트 강화: 5-7줄 설명 및 영어 번역 섹션 강제
-        system_prompt = f"""You are {m_info['name']}. Respond ONLY in {selected_lang}.
-        Provide 5-7 lines of professional music advice. 
-        [TAB] Section: Detailed chords/tabs.
-        [TRANSLATION] English translation of your advice.
-        [MUSIC] English prompt: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}"""
+        # [정교화] 드럼 비트 + 타브 솔로 포함 프롬프트
+        system_prompt = f"""You are the rock star '{m_info['name']}'. Respond ONLY in {selected_lang}.
+        Persona: {persona}.
+
+        CRITICAL RULES:
+        1. Advice: 5-7 deep sentences in {selected_lang}. Speak according to your persona.
+        2. [TAB] SECTION: Must be highly detailed. 
+           - Include Drum patterns (e.g., H|x-x-x-x| S|--o---o-|).
+           - Include Guitar Solo Tab with fret numbers (e.g., e|--12-15-12--|).
+           - Provide precise Chord fingerings.
+        3. [TRANSLATION]: English translation of your advice.
+        4. [MUSIC]: English prompt: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}
+        """
         
         ai_text_raw = ""
         groq_client = ModelManager.get_groq()
@@ -81,35 +101,32 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
         music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
         trans_match = re.search(r'\[TRANSLATION\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
         
-        tab_display = tab_match.group(1).strip() if tab_match else "No Tab Data"
-        translation = trans_match.group(1).strip() if trans_match else ""
-        
-        # TTS용 클린 텍스트 (번역문/음악프롬프트/악보 제외)
-        clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL | re.IGNORECASE)
-        clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text, flags=re.IGNORECASE)
-        clean_text = re.sub(r'\[TRANSLATION\].*?(\[|$)', '', clean_text, flags=re.DOTALL | re.IGNORECASE).strip()
+        tab_display = tab_match.group(1).strip() if tab_match else "No Detailed Tab Generated"
+        eng_translation = trans_match.group(1).strip() if trans_match else ""
+
+        # [수정] TTS 음성에서 번역문 완벽 제외 로직
+        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'[\*\#\-\_\~\|]', '', clean_text)
+        tts_text = re.sub(r'[\*\#\-\_\~\|]', '', speech_text)
         communicate = edge_tts.Communicate(tts_text, m_info["voice"])
         await communicate.save(voice_path)
 
-        # MusicGen (12초)
+        # MusicGen 연주 생성 (12초)
         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)
         scipy.io.wavfile.write(music_path, music_output["sampling_rate"], audio_int16)
 
-        return clean_text, voice_path, music_path, tab_display, translation
+        return speech_text, voice_path, music_path, tab_display, eng_translation
 
     except Exception as e:
-        print(f"Error: {str(e)}")
         return f"Error: {str(e)}", None, None, "No Tab", "Error occurred"
 
 with gr.Blocks() as demo:
-    # 인풋 6개 (script.js의 predict 배열과 일치)
     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)]
     btn = gr.Button("API", visible=False)