app.py

import gradio as gr
import numpy as np
import librosa
from transformers import pipeline
from datetime import datetime
import os
import requests

# 환경변수에서 토큰 가져오기
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
if not HF_API_TOKEN:
    raise ValueError("HF_API_TOKEN not found in environment variables")

# Inference API 설정
API_URL = "https://api-inference.huggingface.co/models/stabilityai/stable-diffusion-xl-base-1.0"
headers = {"Authorization": f"Bearer {HF_API_TOKEN}"}

# AI 모델 초기화
speech_recognizer = pipeline(
    "automatic-speech-recognition",
    model="kresnik/wav2vec2-large-xlsr-korean"
)
emotion_classifier = pipeline(
    "audio-classification",
    model="MIT/ast-finetuned-speech-commands-v2"
)
text_analyzer = pipeline(
    "sentiment-analysis",
    model="nlptown/bert-base-multilingual-uncased-sentiment"
)
korean_sentiment = pipeline(
    "text-classification",
    model="searle-j/korean_sentiment_analysis"
)

# 유틸리티 함수들
def map_acoustic_to_emotion(features):
    """음향학적 특성을 감정으로 매핑"""
    intensity = features["energy"] * 100
    
    if features["energy"] > 0.7:
        if features["tempo"] > 120:
            emotion = "기쁨/흥분"
        else:
            emotion = "분노/강조"
    elif features["pitch"] > 0.6:
        emotion = "놀람/관심"
    elif features["energy"] < 0.3:
        emotion = "슬픔/우울"
    else:
        emotion = "평온/중립"
    
    return {
        "emotion": emotion,
        "intensity": intensity,
        "features": features
    }

def generate_detailed_prompt(text, voice_emotion, text_sentiment, acoustic_features):
    """프롬프트 생성"""
    emotion_colors = {
        "기쁨/흥분": "밝은 노랑과 주황색",
        "분노/강조": "강렬한 빨강과 검정",
        "놀람/관심": "선명한 파랑과 보라",
        "슬픔/우울": "어두운 파랑과 회색",
        "평온/중립": "부드러운 초록과 베이지"
    }
    
    visual_elements = {
        "high_energy": "역동적인 붓질과 강한 대비",
        "medium_energy": "균형잡힌 구도와 자연스러운 흐름",
        "low_energy": "부드러운 그라데이션과 차분한 톤"
    }
    
    energy_level = "medium_energy"
    if acoustic_features["energy"] > 0.7:
        energy_level = "high_energy"
    elif acoustic_features["energy"] < 0.3:
        energy_level = "low_energy"
    
    prompt = f"한국 전통 민화 스타일의 추상화, {emotion_colors.get(voice_emotion['emotion'], '자연스러운 색상')} 기반. "
    prompt += f"{visual_elements[energy_level]}를 통해 감정의 깊이를 표현. "
    prompt += f"음성의 {voice_emotion['emotion']} 감정과 텍스트의 {text_sentiment['label']} 감정이 조화를 이루며, "
    prompt += f"목소리의 특징(강도:{voice_emotion['intensity']:.1f})을 화면의 동적인 요소로 표현. "
    prompt += f"발화 내용 '{text}'의 의미를 은유적 이미지로 담아내기."
    
    return prompt

def generate_image_from_prompt(prompt):
    """이미지 생성"""
    print(f"Generating image with prompt: {prompt}")
    try:
        if not prompt:
            return None
        
        response = requests.post(
            API_URL,
            headers=headers,
            json={
                "inputs": prompt,
                "parameters": {
                    "negative_prompt": "ugly, blurry, poor quality, distorted",
                    "num_inference_steps": 30,
                    "guidance_scale": 7.5
                }
            }
        )
        
        if response.status_code == 200:
            return response.content
        else:
            print(f"Error: {response.status_code}")
            print(f"Response: {response.text}")
            return None
    except Exception as e:
        print(f"Error generating image: {str(e)}")
        return None

def create_interface():
    with gr.Blocks(theme=gr.themes.Soft()) as app:
        # 상태 관리
        state = gr.State({
            "user_name": "",
            "reflections": [],
            "voice_analysis": None,
            "final_prompt": ""
        })

        # 헤더
        header = gr.Markdown("# 디지털 굿판")
        user_display = gr.Markdown("")

        with gr.Tabs() as tabs:
            # 입장
            with gr.Tab("입장"):
                gr.Markdown("""# 디지털 굿판에 오신 것을 환영합니다""")
                name_input = gr.Textbox(label="이름을 알려주세요")
                start_btn = gr.Button("여정 시작하기")

            # 청신
            with gr.Tab("청신"):
                with gr.Row():
                    audio_path = os.path.abspath(os.path.join("assets", "main_music.mp3"))
                    audio = gr.Audio(
                        value=audio_path,
                        type="filepath",
                        label="온천천의 소리",
                        interactive=False,
                        autoplay=True
                    )
                    with gr.Column():
                        reflection_input = gr.Textbox(
                            label="현재 순간의 감상을 적어주세요",
                            lines=3
                        )
                        save_btn = gr.Button("감상 저장하기")
                        reflections_display = gr.Dataframe(
                            headers=["시간", "감상", "감정 분석"],
                            label="기록된 감상들"
                        )

            # 기원
            with gr.Tab("기원"):
                gr.Markdown("## 기원 - 목소리로 전하기")
                with gr.Row():
                    with gr.Column():
                        voice_input = gr.Audio(
                            label="나누고 싶은 이야기를 들려주세요",
                            sources=["microphone"],
                            type="filepath",
                            interactive=True
                        )
                        clear_btn = gr.Button("녹음 지우기")
                    
                    with gr.Column():
                        transcribed_text = gr.Textbox(
                            label="인식된 텍스트",
                            interactive=False
                        )
                        voice_emotion = gr.Textbox(
                            label="음성 감정 분석",
                            interactive=False
                        )
                        text_emotion = gr.Textbox(
                            label="텍스트 감정 분석",
                            interactive=False
                        )
                        analyze_btn = gr.Button("분석하기")

            # 송신
            with gr.Tab("송신"):
                gr.Markdown("## 송신 - 시각화 결과")
                with gr.Column():
                    final_prompt = gr.Textbox(
                        label="생성된 프롬프트",
                        interactive=False,
                        lines=3
                    )
                    generate_btn = gr.Button("이미지 생성하기")
                    result_image = gr.Image(
                        label="생성된 이미지",
                        type="pil"
                    )

        # 인터페이스 함수들
        def start_journey(name):
            """여정 시작"""
            return f"# 환영합니다, {name}님의 디지털 굿판", gr.update(selected="청신")

        def clear_voice_input():
            """음성 입력 초기화"""
            return None

        def analyze_voice(audio_path, state):
            """음성 분석"""
            if audio_path is None:
                return state, "음성을 먼저 녹음해주세요.", "", "", ""
            
            try:
                y, sr = librosa.load(audio_path, sr=16000)
                
                acoustic_features = {
                    "energy": float(np.mean(librosa.feature.rms(y=y))),
                    "tempo": float(librosa.beat.tempo(y)[0]),
                    "pitch": float(np.mean(librosa.feature.zero_crossing_rate(y))),
                    "volume": float(np.mean(np.abs(y)))
                }

                voice_emotion = map_acoustic_to_emotion(acoustic_features)
                transcription = speech_recognizer(y)
                text = transcription["text"]
                text_sentiment = korean_sentiment(text)[0]
                
                voice_result = f"음성 감정: {voice_emotion['emotion']} (강도: {voice_emotion['intensity']:.2f})"
                text_result = f"텍스트 감정: {text_sentiment['label']} ({text_sentiment['score']:.2f})"
                
                prompt = generate_detailed_prompt(text, voice_emotion, text_sentiment, acoustic_features)
                
                return state, text, voice_result, text_result, prompt
            except Exception as e:
                return state, f"오류 발생: {str(e)}", "", "", ""

        def save_reflection(text, state):
            """감상 저장"""
            if not text.strip():
                return state, state["reflections"]
            
            current_time = datetime.now().strftime("%H:%M:%S")
            sentiment = text_analyzer(text)[0]
            new_reflection = [current_time, text, f"{sentiment['label']} ({sentiment['score']:.2f})"]
            
            if "reflections" not in state:
                state["reflections"] = []
            
            state["reflections"].append(new_reflection)
            return state, state["reflections"]

        # 이벤트 연결
        start_btn.click(
            fn=lambda name: (f"# 환영합니다, {name}님의 디지털 굿판", gr.update(selected="청신")),
            inputs=[name_input],
            outputs=[user_display, tabs]
        )

        save_btn.click(
            fn=save_reflection,
            inputs=[reflection_input, state],
            outputs=[state, reflections_display]
        )

        clear_btn.click(
            fn=clear_voice_input,
            inputs=[],
            outputs=[voice_input]
        )

        analyze_btn.click(
            fn=analyze_voice,
            inputs=[voice_input, state],
            outputs=[state, transcribed_text, voice_emotion, text_emotion, final_prompt]
        )

        generate_btn.click(
            fn=generate_image_from_prompt,
            inputs=[final_prompt],
            outputs=[result_image]
        )

        return app

if __name__ == "__main__":
    demo = create_interface()
    demo.launch(debug=True)