app.py

import os, re, tempfile, traceback
import numpy as np
import pandas as pd
import torch
import soundfile as sf
import gradio as gr
from faster_whisper import WhisperModel

# =========================
# إعدادات عامة
# =========================
FORCE_WHISPER_NAME = "large-v3"
FORCE_COMPUTE_TYPE = "int8"
ASR_OPTS = dict(
    word_timestamps=True,
    vad_filter=True,
    vad_parameters={"min_silence_duration_ms": 200},
    beam_size=5,
    best_of=5,
    temperature=0.0,
)

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[INIT] Using device: {DEVICE}", flush=True)

# =========================
# تحميل نموذج Whisper
# =========================
_WHISPER = None
def load_whisper_model():
    global _WHISPER
    if _WHISPER is None:
        _WHISPER = WhisperModel(FORCE_WHISPER_NAME,
                                device=DEVICE,
                                compute_type=FORCE_COMPUTE_TYPE)
        print(f"[LOAD] Whisper model: {FORCE_WHISPER_NAME} ({FORCE_COMPUTE_TYPE})", flush=True)
    return _WHISPER

whisper = load_whisper_model()

# =========================
# دوال مساعدة
# =========================
def ensure_audio_path(audio):
    """يتأكد إنو الإدخال الصوتي محفوظ مؤقتًا كملف wav"""
    if isinstance(audio, str):
        if not os.path.exists(audio):
            raise FileNotFoundError(f"Audio file not found: {audio}")
        return audio
    if isinstance(audio, tuple) and len(audio) == 2:
        data, sr = audio
        if isinstance(data, np.ndarray):
            tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
            sf.write(tmp.name, data, sr)
            return tmp.name
    raise ValueError("Unsupported audio input format")

def normalize_ar_orth(text: str) -> str:
    text = re.sub(r"[ًٌٍَُِّْـ]", "", text)
    text = re.sub(r"[“”\"',:؛؟.!()\[\]{}،\-–—_]", " ", text)
    text = re.sub(r"\s+", " ", text).strip()
    return text

def clean_ar_token(t: str) -> str:
    t = t.strip()
    t = re.sub(r'^[^\w\u0600-\u06FF]+|[^\w\u0600-\u06FF]+$', '', t)
    t = normalize_ar_orth(t)
    return t

def extract_word_conf_table(segments):
    """يبني جدول الكلمات مع الزمن والثقة"""
    rows = []
    for seg in segments:
        for w in (seg.words or []):
            rows.append({
                "seg_start": float(seg.start),
                "seg_end": float(seg.end),
                "word_start": float(w.start),
                "word_end": float(w.end),
                "word": clean_ar_token(w.word),
                "prob": float(w.probability),
            })
    return pd.DataFrame(rows)

# =========================
# الدالة الأساسية
# =========================
def transcribe_audio(audio):
    try:
        audio_path = ensure_audio_path(audio)
        segments, info = whisper.transcribe(audio_path, **ASR_OPTS)
        full_text = " ".join([seg.text for seg in segments])

        df_words = extract_word_conf_table(segments)
        if df_words.empty:
            # نرجع جدول فاضي بدل النص حتى ما يصير error
            df_words = pd.DataFrame(columns=["seg_start", "seg_end", "word_start", "word_end", "word", "prob"])

        return full_text, df_words

    except Exception as e:
        return f"❌ حدث خطأ أثناء المعالجة:\n{str(e)}\n\n{traceback.format_exc()}", pd.DataFrame()

# =========================
# واجهة Gradio
# =========================
demo = gr.Interface(
    fn=transcribe_audio,
    inputs=gr.Audio(sources=["upload", "microphone"], type="filepath", label="🎙️ ملف الصوت أو التسجيل"),
    outputs=[
        gr.Textbox(label="النص الناتج"),
        gr.Dataframe(label="تفاصيل الكلمات (زمن البداية، النهاية، الثقة)")
    ],
    title="Arabic Whisper ASR",
    description="حوّل الكلام العربي إلى نص باستخدام Faster-Whisper (large-v3).",
    theme="soft",
    allow_flagging="never"
)

if __name__ == "__main__":
    demo.launch(share=True, show_error=True)