Update app.py

app.py

@@ -4,6 +4,7 @@ import gradio as gr
 from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
 from groq import Groq
 
+# 모델 관리 싱글톤
 class ModelManager:
     _llm_pipeline = None
     _music_pipeline = None
@@ -31,7 +32,7 @@ class ModelManager:
             cls._music_pipeline = pipeline("text-to-audio", "facebook/musicgen-small", device="cpu")
         return cls._music_pipeline
 
-# 언어별 전담 멤버 설정
+# 언어별 멤버/보이스 매핑
 LANG_MEMBER_MAP = {
     "Korean": {"name": "서윤 (Korea)", "voice": "ko-KR-SunHiNeural"},
     "English": {"name": "Chloe (USA)", "voice": "en-US-AriaNeural"},
@@ -48,12 +49,14 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
     voice_path = f"/tmp/v_{req_id}.mp3"
     music_path = f"/tmp/m_{req_id}.wav"
     
-    member_info = LANG_MEMBER_MAP.get(selected_lang, LANG_MEMBER_MAP["Korean"])
+    m_info = LANG_MEMBER_MAP.get(selected_lang, LANG_MEMBER_MAP["Korean"])
     
-    system_prompt = f"""You are {member_info['name']}. Respond ONLY in {selected_lang}.
-    Provide 5-7 lines of professional music advice. 
+    # [TRANSLATION] 섹션을 추가하도록 프롬프트 수정
+    system_prompt = f"""You are {m_info['name']}. Respond ONLY in {selected_lang}.
+    Provide 5-7 sentences of music advice. 
+    At the end, add [TRANSLATION] followed by the English translation of your advice.
     [TAB] Section: Detailed chords/tabs.
-    [MUSIC] Section: English prompt reflecting: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}"""
+    [MUSIC] Section: English prompt for MusicGen: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}"""
     
     ai_text_raw = ""
     groq_client = ModelManager.get_groq()
@@ -71,15 +74,23 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
         out = qwen(f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{user_input}<|im_end|>\nassistant\n", max_new_tokens=1024)
         ai_text_raw = out[0]['generated_text'].split("assistant\n")[-1]
 
+    # 데이터 파싱 (본문, 번역, 악보, 음악)
     tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
     music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
-    tab_display = tab_match.group(1).strip() if tab_match else "No Score Data"
+    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 Data"
+    translation = trans_match.group(1).strip() if trans_match else ""
+    
     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).strip()
+    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()
     
-    communicate = edge_tts.Communicate(re.sub(r'[\*\#\-\_\~\|]', '', clean_text), member_info["voice"])
+    # TTS 생성 (네이티브 언어로)
+    communicate = edge_tts.Communicate(re.sub(r'[\*\#\-\_\~\|]', '', clean_text), m_info["voice"])
     await communicate.save(voice_path)
 
+    # MusicGen (12초)
     music_gen = ModelManager.get_music()
     music_p = music_match.group(1).strip() if music_match else "rock"
     music_output = music_gen(music_p, forward_params={"max_new_tokens": 512})
@@ -87,11 +98,12 @@ async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, cho
     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
+    # 리턴 시 번역 포함 (clean_text, voice_path, music_path, tab_display, translation)
+    return clean_text, voice_path, music_path, tab_display, translation
 
 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)]
+    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")