Update app.py

app.py

@@ -2,133 +2,125 @@ import gradio as gr
 import requests
 import edge_tts
 import os
-import json
+import re
+import torch
+import scipy.io.wavfile
+import numpy as np
+from transformers import pipeline
 
-# 1. API 키 설정
+# 1. API 키 (환경변수 설정 필수!)
 GENAI_KEY = os.getenv("GEMINI_KEY")
 
-# ======================================================
-# 🕵️‍♂️ [핵심] 구글 서버에서 '진짜 모델 이름' 알아오기
-# ======================================================
+# 2. [작곡가] MusicGen 모델 로드
+print("⏳ 작곡가(MusicGen) 섭외 중... (최대 1~2분 소요)")
+try:
+    # facebook/musicgen-small : 가볍고 빠른 모델
+    music_synthesiser = pipeline("text-to-audio", "facebook/musicgen-small")
+    print("✅ 작곡가 섭외 완료! Start Jamming!")
+except Exception as e:
+    print(f"⚠️ 작곡 모델 로드 실패: {e}")
+    music_synthesiser = None
+
+# 3. Gemini 모델 찾기 헬퍼
 def find_working_model():
-    # 모델 리스트 조회 API 호출
     url = f"https://generativelanguage.googleapis.com/v1beta/models?key={GENAI_KEY}"
-    response = requests.get(url)
-    
-    if response.status_code != 200:
-        print(f"❌ 모델 리스트 조회 실패: {response.text}")
-        return "gemini-pro" # 실패하면 그냥 깡통 모델로 잡기
-
-    data = response.json()
-    
-    # 1. 'generateContent' 기능이 있는 모델만 필터링
-    candidates = []
-    if 'models' in data:
-        for m in data['models']:
-            if 'supportedGenerationMethods' in m and 'generateContent' in m['supportedGenerationMethods']:
-                # 이름에서 'models/' 부분 떼고 저장 (예: models/gemini-pro -> gemini-pro)
-                clean_name = m['name'].replace("models/", "")
-                candidates.append(clean_name)
-    
-    print(f"📋 사용 가능 모델 목록: {candidates}")
-
-    # 2. 우선순위대로 선택 (플래시 -> 프로)
-    # 리더님 키로 'gemini-1.5-flash'가 안되면 'gemini-1.0-pro'라도 잡습니다.
-    preferred_order = [
-        "gemini-1.5-flash-latest",
-        "gemini-1.5-flash",
-        "gemini-1.5-flash-001",
-        "gemini-1.5-pro",
-        "gemini-1.0-pro",
-        "gemini-pro"
-    ]
-    
-    for pref in preferred_order:
-        if pref in candidates:
-            print(f"✅ 결정된 모델: {pref}")
-            return pref
-            
-    # 선호하는 게 없으면 목록의 첫 번째 놈 잡기
-    if candidates:
-        print(f"⚠️ 선호 모델 없음. 대체 모델 사용: {candidates[0]}")
-        return candidates[0]
-        
-    return "gemini-pro" # 진짜 아무것도 없으면...
-
-# ======================================================
-# 3. 직접 통신 함수 (모델명을 동적으로 넣음)
-# ======================================================
-def call_gemini_direct(prompt, model_name):
-    # 위에서 찾은 모델명을 URL에 끼워넣기
-    url = f"https://generativelanguage.googleapis.com/v1beta/models/{model_name}:generateContent?key={GENAI_KEY}"
-    
-    headers = {'Content-Type': 'application/json'}
-    data = {
-        "contents": [{
-            "parts": [{"text": prompt}]
-        }]
-    }
-    
-    response = requests.post(url, headers=headers, json=data)
-    
-    if response.status_code == 200:
-        return response.json()['candidates'][0]['content']['parts'][0]['text']
-    else:
-        # 에러 나면 에러 메시지 반환
-        raise Exception(f"Google API Error ({model_name}): {response.text}")
+    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"]
+        for p in preferred:
+            if p in candidates: return p
+        return candidates[0] if candidates else "gemini-pro"
+    except:
+        return "gemini-pro"
 
+ACTIVE_MODEL = find_working_model()
 
-# 4. 멤버 데이터 (그대로)
+# 4. 멤버 페르소나 정의
 MEMBERS = {
-    "서윤 (Korea)": { "role": "메인 보컬", "voice": "ko-KR-SunHiNeural", "prompt": "당신은 록밴드 'Error 404'의 메인보컬 '서윤'입니다. 한국어(Korean)로 대답하세요. 성격: 시니컬, 츤데레. 말투: '야', '있잖아' 사용." },
-    "Chloe (USA)": { "role": "리드 기타", "voice": "en-US-AriaNeural", "prompt": "You are 'Chloe'. Respond in English. Personality: Cool rockstar." },
-    "Beatrice (Brazil)": { "role": "드럼", "voice": "pt-BR-FranciscaNeural", "prompt": "You are 'Beatrice'. Respond in English/Portuguese. Personality: Passionate." },
-    "Naomi (Japan)": { "role": "리듬 기타", "voice": "ja-JP-NanamiNeural", "prompt": "You are 'Naomi'. Respond in Japanese. Personality: Calm." },
-    "Elena (Spain)": { "role": "베이스", "voice": "es-ES-ElviraNeural", "prompt": "You are 'Elena'. Respond in Spanish/English. Personality: Mature." },
-    "Amira (Egypt)": { "role": "키보드", "voice": "ar-EG-SalmaNeural", "prompt": "You are 'Amira'. Respond in Arabic/English. Personality: Mystical." },
-    "Liwei (China)": { "role": "DJ / FX", "voice": "zh-CN-XiaoxiaoNeural", "prompt": "You are 'Liwei'. Respond in Chinese/English. Personality: Trendy." },
-    "Sophie (France)": { "role": "바이올린", "voice": "fr-FR-DeniseNeural", "prompt": "You are 'Sophie'. Respond in French/English. Personality: Elegant." }
+    "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." }
 }
 
 async def band_consulting(user_input, member_name):
     try:
         if member_name not in MEMBERS:
-            return f"Error: 이름 불일치! ('{member_name}')", None
+            return "Error: Unknown Member", None, None
 
         member = MEMBERS[member_name]
-        full_prompt = f"{member['prompt']}\n\nUser's Error/Worry: {user_input}\nResponse:"
         
-        try:
-            # 1. 살아있는 모델 찾기 (매번 찾으면 느리니까 캐싱하면 좋지만, 확실함을 위해 매번 확인)
-            active_model = find_working_model()
-            
-            # 2. 그 모델로 요청 보내기
-            ai_text = call_gemini_direct(full_prompt, active_model)
+        # ★ 시스템 프롬프트: 작곡 명령 [MUSIC:...] 을 유도
+        system_instruction = f"""
+        {member['prompt']}
+        
+        [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:...].
+        """
+        
+        # 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
             
-        except Exception as e:
-            return f"Error (Gemini): {str(e)}", None
+        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)
+        
+        if match:
+            music_prompt = match.group(1).strip()
+            clean_text = ai_text_raw.replace(match.group(0), "").strip()
+            print(f"🎵 작곡 요청: {music_prompt}")
 
-        try:
-            output_file = f"/tmp/{member_name}_reply.mp3"
-            communicate = edge_tts.Communicate(ai_text, member['voice'])
-            await communicate.save(output_file)
-        except Exception as e:
-            return f"Error (TTS): {str(e)}", None
+        # TTS (목소리) 생성
+        voice_path = f"/tmp/{member_name}_voice.mp3"
+        await edge_tts.Communicate(clean_text, member['voice']).save(voice_path)
         
-        return ai_text, output_file
+        # MusicGen (음악) 생성
+        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}")
+
+        return clean_text, voice_path, music_path
 
     except Exception as e:
-        return f"System Error: {str(e)}", None
+        return f"System Error: {e}", None, None
 
-# UI
+# Gradio 앱 실행
 with gr.Blocks() as demo:
-    gr.Markdown("# Error 404 Band API Server (Auto-Discovery Mode)")
     with gr.Row():
-        inp_text = gr.Textbox(label="고민 입력")
-        inp_member = gr.Dropdown(choices=list(MEMBERS.keys()), label="멤버 선택", value="서윤 (Korea)")
-    out_text = gr.Textbox(label="조언")
-    out_audio = gr.Audio(label="음성")
-    btn = gr.Button("상담 요청")
-    btn.click(fn=band_consulting, inputs=[inp_text, inp_member], outputs=[out_text, out_audio], api_name="predict")
+        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")
 
 demo.launch()