app.py

import gradio as gr
import torch
from transformers import pipeline, AutoModelForSeq2SeqLM, AutoTokenizer
from moviepy.video.io.VideoFileClip import VideoFileClip
import librosa
import soundfile as sf
import numpy as np
import re
import os

# ======================
# 1. الإعدادات الأساسية
# ======================
device = "cuda" if torch.cuda.is_available() else "cpu"

# ======================
# 2. تحميل النماذج
# ======================
pipe = pipeline("automatic-speech-recognition", model="openai/whisper-medium", device=0 if device=="cuda" else -1)
bart_model = AutoModelForSeq2SeqLM.from_pretrained("ahmedabdo/arabic-summarizer-bart")
bart_tokenizer = AutoTokenizer.from_pretrained("ahmedabdo/arabic-summarizer-bart")

# ======================
# 3. الدوال المساعدة
# ======================
def clean_text(text):
    return re.sub(r'\s+', ' ', text).strip()

def convert_audio_to_text(uploaded_file):
    try:
        if not uploaded_file:
            return "⛔ الرجاء رفع ملف أولاً"

        input_path = uploaded_file if isinstance(uploaded_file, str) else uploaded_file.name
        output_path = "/tmp/processed.wav"

        # معالجة ملفات الفيديو
        if input_path.split('.')[-1].lower() in ['mp4', 'avi', 'mov', 'mkv']:
            video = VideoFileClip(input_path)
            if video.audio:
                video.audio.write_audiofile(output_path, codec='pcm_s16le')
            else:
                return "⛔ لا يوجد صوت في الفيديو!"
        else:
            output_path = input_path

        audio, rate = librosa.load(output_path, sr=16000)
        transcripts = []
        # تقسيم الصوت إلى مقاطع للتعامل مع الملفات الكبيرة
        for start in np.arange(0, len(audio)/rate, 30):
            end = min(start + 30, len(audio)/rate)
            segment = audio[int(start*rate):int(end*rate)]
            sf.write(f"/tmp/segment_{int(start)}.wav", segment, rate)
            transcripts.append(pipe(f"/tmp/segment_{int(start)}.wav")["text"])
        return " ".join(transcripts)
    except Exception as e:
        return f"⛔ خطأ: {str(e)}"

def summarize_text(text):
    cleaned_text = clean_text(text)
    inputs = bart_tokenizer(cleaned_text, return_tensors="pt", max_length=1024, truncation=True).to(device)
    summary_ids = bart_model.generate(
        inputs.input_ids,
        max_length=150,
        num_beams=4,
        early_stopping=True
    )
    return bart_tokenizer.decode(summary_ids[0], skip_special_tokens=True)

# ======================
# 4. المثال التجريبي
# ======================
EXAMPLE_AUDIO_PATH = "AUDIO-2025-02-24-22-10-37.mp3"

def process_example_audio():
    if not os.path.exists(EXAMPLE_AUDIO_PATH):
        return "⛔ الملف التجريبي غير موجود"
    return convert_audio_to_text(EXAMPLE_AUDIO_PATH)

# ======================
# 5. واجهة المستخدم
# ======================
with gr.Blocks() as demo:
    gr.Markdown("## 🎤 استخراج النص وتلخيصه")
    
    gr.Markdown("### 🔊 مثال تجريبي")
    gr.Audio(EXAMPLE_AUDIO_PATH)
    example_btn = gr.Button("تجربة المثال ⚡", elem_classes="custom-button")
    
    file_input = gr.File(file_types=[".wav", ".mp3", ".mp4"])
    extract_btn = gr.Button("استخراج النص", elem_classes="custom-button")
    extracted_text = gr.Textbox(label="📝 النص المستخرج", lines=8)
    summarize_btn = gr.Button("تلخيص النص", elem_classes="custom-button")
    summary_output = gr.Textbox(label="📌 الملخص", lines=6)
    
    extract_btn.click(convert_audio_to_text, inputs=file_input, outputs=extracted_text)
    summarize_btn.click(summarize_text, inputs=extracted_text, outputs=summary_output)
    example_btn.click(process_example_audio, outputs=extracted_text)

if __name__ == "__main__":
    demo.launch()