Update app.py

app.py

@@ -3,95 +3,76 @@ import requests
 import edge_tts
 import os
 import re
-import torch
 import scipy.io.wavfile
-import numpy as np
 from transformers import pipeline
 
 GENAI_KEY = os.getenv("GEMINI_KEY")
-
 try:
-    # 음질 보정 모델 로드
     music_synthesiser = pipeline("text-to-audio", "facebook/musicgen-small")
-except Exception as e:
+except:
     music_synthesiser = None
 
-def find_working_model():
-    url = f"https://generativelanguage.googleapis.com/v1beta/models?key={GENAI_KEY}"
-    try:
-        response = requests.get(url)
-        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-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()
-
+# 멤버 설정 (목소리 매칭용)
 MEMBERS = {
-    "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." }
+    "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"
 }
 
-async def band_consulting(user_input, member_name, consult_category):
+async def band_consulting(user_input, member_name, consult_lang_code):
     try:
-        member = MEMBERS.get(member_name, MEMBERS["서윤 (Korea)"])
-        # 리더님 요청: 영어 상담 시 무조건 영어만 사용
-        requested_lang = "English" if "English" in consult_category else "Korean"
-
+        # 선택된 언어 코드(ko, en, ja 등)에 따라 답변 언어 강제
+        lang_map = {"ko": "Korean", "en": "English", "ja": "Japanese", "pt": "Portuguese", "es": "Spanish", "ar": "Arabic", "zh": "Chinese", "fr": "French"}
+        target_lang = lang_map.get(consult_lang_code, "Korean")
+        
         system_instruction = f"""
-        {member['prompt']}
-        [LANGUAGE RULE] Respond 100% in {requested_lang}. NEVER use other languages.
-        [RESPONSE STYLE] 
-        1. Keep your verbal response (main text) VERY CONCISE, under 5 short lines. 
-        2. Put ALL technical details, scales, and long musical advice inside the [TAB] section.
-        3. Start with a charismatic rockstar greeting.
-
-        After advice, you MUST include:
-        1. [TAB]: Detailed musical explanation + Chord/Tab progression.
-        2. [MUSIC]: English music prompt (high quality, studio recording).
+        당신은 밴드 멤버 '{member_name}'입니다.
+        [규칙 1] 무조건 '{target_lang}'로만 답변하세요.
+        [규칙 2] 음성 답변용 본문은 5줄 이내로 짧게 요약하세요.
+        [규칙 3] 모든 상세 설명과 악보는 [TAB] 섹션에 넣으세요.
+        [규칙 4] 음악 생성용 영어 프롬프트는 [MUSIC: 프롬프트] 형식으로 끝에 포함하세요.
         """
         
-        url = f"https://generativelanguage.googleapis.com/v1beta/models/{ACTIVE_MODEL}:generateContent?key={GENAI_KEY}"
+        # Gemini API 호출 (모델명은 환경에 맞게 자동 조절된다고 가정)
+        url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key={GENAI_KEY}"
         payload = {"contents": [{"parts": [{"text": f"{system_instruction}\n질문: {user_input}"}]}]}
-        res = requests.post(url, json=payload, headers={'Content-Type': 'application/json'})
-        ai_text_raw = res.json()['candidates'][0]['content']['parts'][0]['text']
+        res = requests.post(url, json=payload).json()
+        ai_text_raw = res['candidates'][0]['content']['parts'][0]['text']
         
+        # 데이터 파싱
         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 "상세 악보 없음"
         
+        # TTS용 클린 텍스트 (악보/프롬프트 제거 + 5줄 컷)
         clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL)
         clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text).strip()
-        tts_text = re.sub(r'[\*\#\-\_\~]', '', clean_text) 
+        clean_text = "\n".join(clean_text.split("\n")[:5])
 
-        voice_path = f"/tmp/v_{member_name.replace(' ', '_')}.mp3"
-        await edge_tts.Communicate(tts_text, member['voice']).save(voice_path)
+        # 목소리 생성
+        voice_path = f"/tmp/v_{member_name.split()[0]}.mp3"
+        await edge_tts.Communicate(clean_text, MEMBERS.get(member_name, "ko-KR-SunHiNeural")).save(voice_path)
         
+        # 음악 생성
         music_path = None
         if music_match and music_synthesiser:
             p = music_match.group(1).strip() + ", high quality, studio recording"
-            # 리더님 요청: 길이 약 10~12초(512 tokens)
-            music_output = music_synthesiser(p, forward_params={"max_new_tokens": 512, "guidance_scale": 4.5, "do_sample": True})
-            music_path = f"/tmp/m_{member_name.replace(' ', '_')}.wav"
+            music_output = music_synthesiser(p, forward_params={"max_new_tokens": 512, "do_sample": True})
+            music_path = f"/tmp/m_{member_name.split()[0]}.wav"
             scipy.io.wavfile.write(music_path, music_output["sampling_rate"], music_output["audio"][0].T)
 
         return clean_text, voice_path, music_path, tab_display
     except Exception as e:
-        return f"Error: {str(e)}", None, None, "Error"
+        return str(e), None, None, "Error"
 
+# Gradio 인터페이스는 API용으로 최소화
 with gr.Blocks() as demo:
-    with gr.Row():
-        i1 = gr.Textbox(); i2 = gr.Textbox(); i3 = gr.Textbox()
+    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()