Update app.py

app.py

@@ -8,119 +8,93 @@ import scipy.io.wavfile
 import numpy as np
 from transformers import pipeline
 
-# 1. API 키 (환경변수 설정 필수!)
+# API 키 설정
 GENAI_KEY = os.getenv("GEMINI_KEY")
 
-# 2. [작곡가] MusicGen 모델 로드
-print("⏳ 작곡가(MusicGen) 섭외 중... (최대 1~2분 소요)")
+# 1. [작곡가] MusicGen 로드 (음질 보정 설정 준비)
+print("⏳ 작곡가 섭외 중... (음질 보정 모드)")
 try:
-    # facebook/musicgen-small : 가볍고 빠른 모델
     music_synthesiser = pipeline("text-to-audio", "facebook/musicgen-small")
-    print("✅ 작곡가 섭외 완료! Start Jamming!")
+    print("✅ 작곡가 준비 완료!")
 except Exception as e:
-    print(f"⚠️ 작곡 모델 로드 실패: {e}")
+    print(f"⚠️ 모델 로드 실패: {e}")
     music_synthesiser = None
 
-# 3. Gemini 모델 찾기 헬퍼
 def find_working_model():
     url = f"https://generativelanguage.googleapis.com/v1beta/models?key={GENAI_KEY}"
     try:
         response = requests.get(url)
-        if response.status_code != 200: return "gemini-pro"
         data = response.json()
         candidates = [m['name'].replace("models/", "") for m in data.get('models', []) if 'generateContent' in m.get('supportedGenerationMethods', [])]
-        preferred = ["gemini-1.5-flash", "gemini-1.5-flash-latest", "gemini-pro"]
+        preferred = ["gemini-1.5-flash", "gemini-pro"]
         for p in preferred:
             if p in candidates: return p
         return candidates[0] if candidates else "gemini-pro"
-    except:
-        return "gemini-pro"
+    except: return "gemini-pro"
 
 ACTIVE_MODEL = find_working_model()
 
-# 4. 멤버 페르소나 정의
+# 멤버 페르소나 (악보 생성 능력 추가)
 MEMBERS = {
-    "서윤 (Korea)": { "voice": "ko-KR-SunHiNeural", "prompt": "당신은 록밴드 보컬 '서윤'입니다. 한국어(Korean)로 대답하세요. 성격: 시니컬하지만 음악엔 진심. 답변 끝에 작곡 아이디어를 포함하세요." },
-    "Chloe (USA)": { "voice": "en-US-AriaNeural", "prompt": "You are 'Chloe', the lead guitarist. Speak in English. Personality: Cool rockstar. Suggest guitar riffs." },
-    "Beatrice (Brazil)": { "voice": "pt-BR-FranciscaNeural", "prompt": "You are 'Beatrice', the drummer. Speak in English/Portuguese. Focus on rhythm and groove." },
-    "Naomi (Japan)": { "voice": "ja-JP-NanamiNeural", "prompt": "You are 'Naomi'. Speak in Japanese. Focus on melody and chords." },
-    "Elena (Spain)": { "voice": "es-ES-ElviraNeural", "prompt": "You are 'Elena'. Speak in Spanish/English. Focus on bass lines." },
-    "Amira (Egypt)": { "voice": "ar-EG-SalmaNeural", "prompt": "You are 'Amira'. Speak in Arabic/English. Focus on atmosphere." },
-    "Liwei (China)": { "voice": "zh-CN-XiaoxiaoNeural", "prompt": "You are 'Liwei'. Speak in Chinese/English. Focus on FX and sound design." },
-    "Sophie (France)": { "voice": "fr-FR-DeniseNeural", "prompt": "You are 'Sophie'. Speak in French/English. Focus on emotional melody." }
+    "서윤 (Korea)": { "voice": "ko-KR-SunHiNeural", "prompt": "당신은 보컬 '서윤'입니다. 시니컬한 말투. 상담 분야에 맞춰 음악적 조언과 코드 진행을 알려주세요." },
+    "Chloe (USA)": { "voice": "en-US-AriaNeural", "prompt": "You are 'Chloe', the guitarist. Speak in English. Focus on guitar tabs and tones." },
+    "Beatrice (Brazil)": { "voice": "pt-BR-FranciscaNeural", "prompt": "You are 'Beatrice', the drummer. Focus on rhythm patterns and groovy beats." }
 }
 
-async def band_consulting(user_input, member_name):
+async def band_consulting(user_input, member_name, consult_category):
     try:
-        if member_name not in MEMBERS:
-            return "Error: Unknown Member", None, None
-
-        member = MEMBERS[member_name]
+        member = MEMBERS.get(member_name, MEMBERS["서yun (Korea)"])
         
-        # ★ 시스템 프롬프트: 작곡 명령 [MUSIC:...] 을 유도
+        # 시스템 프롬프트: 악보와 음악 생성을 구체화
         system_instruction = f"""
         {member['prompt']}
+        [상담 분야: {consult_category}]
         
-        [INSTRUCTION]
-        1. Give advice based on the user's input.
-        2. IF appropriate (user asks for music, sample, or feeling), create a short music sample.
-        3. To create music, write a hidden command at the END of your response like this:
-           [MUSIC: genre, tempo, instruments, mood]
-        
-        Example: "Let's go with a heavy beat! [MUSIC: heavy metal drum beat, 140bpm, aggressive]"
-        If no music needed, do NOT write [MUSIC:...].
+        당신은 전문가로서 조언을 하고, 마지막에 두 가지를 반드시 포함하세요:
+        1. [TAB]: 음악의 간단한 악보나 코드 진행을 텍스트 형식으로 그리세요. 
+           예: | Am - F - C - G |
+        2. [MUSIC]: 분야에 특화된 고음질 음악 프롬프트를 적으세요. 
+           (High-quality, studio recording, master audio 필수 포함)
         """
         
-        # Gemini 호출
         url = f"https://generativelanguage.googleapis.com/v1beta/models/{ACTIVE_MODEL}:generateContent?key={GENAI_KEY}"
-        payload = {"contents": [{"parts": [{"text": f"{system_instruction}\nUser: {user_input}\nResponse:"}]}]}
-        res = requests.post(url, json=payload, headers={'Content-Type': 'application/json'})
-        
-        if res.status_code != 200:
-            return f"API Error: {res.text}", None, None
-            
+        payload = {"contents": [{"parts": [{"text": f"{system_instruction}\n질문: {user_input}"}]}]}
+        res = requests.post(url, json=payload)
         ai_text_raw = res.json()['candidates'][0]['content']['parts'][0]['text']
         
-        # 음악 명령 추출
-        music_prompt = None
-        clean_text = ai_text_raw
-        match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
+        # 악보([TAB])와 음악([MUSIC]) 추출
+        tab_match = re.search(r'\[TAB\](.*?)\[', ai_text_raw + '[', re.DOTALL)
+        music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
         
-        if match:
-            music_prompt = match.group(1).strip()
-            clean_text = ai_text_raw.replace(match.group(0), "").strip()
-            print(f"🎵 작곡 요청: {music_prompt}")
+        tab_display = tab_match.group(1).strip() if tab_match else "No Score Available"
+        clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL)
+        clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text).strip()
 
-        # TTS (목소리) 생성
-        voice_path = f"/tmp/{member_name}_voice.mp3"
+        # 1. 목소리 생성
+        voice_path = f"/tmp/v_{member_name}.mp3"
         await edge_tts.Communicate(clean_text, member['voice']).save(voice_path)
         
-        # MusicGen (음악) 생성
+        # 2. 음악 생성 (음질 보정 트릭 적용)
         music_path = None
-        if music_prompt and music_synthesiser:
-            try:
-                # max_new_tokens=256 (약 5초) ~ 512 (약 10초)
-                music = music_synthesiser(music_prompt, forward_params={"max_new_tokens": 512})
-                music_path = f"/tmp/{member_name}_music.wav"
-                scipy.io.wavfile.write(music_path, music["sampling_rate"], music["audio"][0].T)
-            except Exception as e:
-                print(f"❌ 작곡 실패: {e}")
+        if music_match and music_synthesiser:
+            p = music_match.group(1).strip() + ", high quality, studio recording, clear sound"
+            # 음질 보정: guidance_scale을 높여 프롬프트 충실도 향상
+            music = music_synthesiser(p, forward_params={"max_new_tokens": 448, "guidance_scale": 4.5, "do_sample": True})
+            music_path = f"/tmp/m_{member_name}.wav"
+            scipy.io.wavfile.write(music_path, music["sampling_rate"], music["audio"][0].T)
 
-        return clean_text, voice_path, music_path
+        return clean_text, voice_path, music_path, tab_display
 
     except Exception as e:
-        return f"System Error: {e}", None, None
+        return f"Error: {e}", None, None, ""
 
-# Gradio 앱 실행
+# Gradio 인터페이스
 with gr.Blocks() as demo:
     with gr.Row():
-        inp = gr.Textbox(label="Input")
-        mem = gr.Textbox(label="Member")
-    out_text = gr.Textbox(label="Advice")
-    out_voice = gr.Audio(label="Voice")
-    out_music = gr.Audio(label="Music Sample")
-    
-    btn = gr.Button("GO")
-    btn.click(band_consulting, [inp, mem], [out_text, out_voice, out_music], api_name="predict")
+        inp = gr.Textbox()
+        mem = gr.Textbox()
+        cat = gr.Textbox() # 분야
+    btn = gr.Button("Predict")
+    btn.click(band_consulting, [inp, mem, cat], [gr.Textbox(), gr.Audio(), gr.Audio(), gr.Textbox()], api_name="predict")
 
 demo.launch()