app.py

import gradio as gr
import yt_dlp
import os
import shutil
import subprocess
from faster_whisper import WhisperModel

# 🔤 Hindi Script Fix
from indic_transliteration import sanscript
from indic_transliteration.sanscript import transliterate

# ===============================
# 1. Whisper Model (Lazy Load)
# ===============================
model = None

def load_model():
    global model
    if model is None:
        print("📥 Loading Whisper Model...")
        model = WhisperModel("base", device="cpu", compute_type="int8")
        print("✅ Model Loaded")
    return model

# ===============================
# 2. FFmpeg Path
# ===============================
def get_ffmpeg():
    return shutil.which("ffmpeg") or "/usr/bin/ffmpeg"

# ===============================
# 3. Video → Audio
# ===============================
def extract_audio(video_path):
    audio_path = "uploaded_audio.wav"
    if os.path.exists(audio_path):
        os.remove(audio_path)

    cmd = [
        get_ffmpeg(),
        "-i", video_path,
        "-vn",
        "-ac", "1",
        "-ar", "16000",
        audio_path,
        "-y"
    ]
    subprocess.run(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
    return audio_path

# ===============================
# 4. Download Audio from URL
# ===============================
def download_audio_from_url(url):
    output = "url_audio"

    ydl_opts = {
        "format": "bestaudio/best",
        "outtmpl": output,
        "postprocessors": [{
            "key": "FFmpegExtractAudio",
            "preferredcodec": "wav",
        }],
        "quiet": True,
        "nocheckcertificate": True,
    }

    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        ydl.download([url])

    return "url_audio.wav"

# ===============================
# 5. Hindi Script Normalizer
# ===============================
def normalize_script(text, lang):
    if lang == "hi":
        try:
            return transliterate(text, sanscript.ARABIC, sanscript.DEVANAGARI)
        except:
            return text
    return text

# ===============================
# 6. Main Transcribe Logic
# ===============================
def transcribe_media(url_input, file_input, language_choice):

    try:
        audio_path = None

        # ---------- FILE ----------
        if file_input:
            ext = os.path.splitext(file_input)[1].lower()
            if ext in [".mp3", ".wav", ".m4a"]:
                audio_path = file_input
            else:
                audio_path = extract_audio(file_input)

        # ---------- URL ----------
        elif url_input and url_input.strip():
            audio_path = download_audio_from_url(url_input)

        else:
            return "⚠️ Please paste a link or upload a file."

        if not os.path.exists(audio_path):
            return "❌ Audio processing failed."

        model = load_model()

        # Language handling
        language = None if language_choice == "Auto Detect" else language_choice

        segments, info = model.transcribe(
            audio_path,
            beam_size=1,
            vad_filter=True,
            language=language
        )

        detected_lang = info.language
        raw_text = " ".join(seg.text for seg in segments)
        final_text = normalize_script(raw_text, detected_lang)

        return f"🌍 Detected Language: {detected_lang}\n\n{final_text.strip()}"

    except Exception as e:
        return f"❌ Error: {str(e)}"

# ===============================
# 7. UI
# ===============================
css = """
.container {max-width: 900px; margin: auto;}
.gr-button-primary {
    background: linear-gradient(90deg,#667eea,#764ba2);
    border: none;
    color: white;
}
"""

with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
    with gr.Column(elem_classes="container"):
        gr.Markdown("## 🚀 Universal Transcript Tool")
        gr.Markdown(
            "Supports **YouTube, TikTok, Instagram, Facebook, Twitter/X**\n\n"
            "Hindi output is always **Devanagari** 🇮🇳"
        )

        with gr.Tabs():
            with gr.TabItem("🔗 Paste Link"):
                url_in = gr.Textbox(label="Video URL")
                btn_url = gr.Button("🎧 Transcribe Link", variant="primary")

            with gr.TabItem("📂 Upload File"):
                file_in = gr.File(
                    label="Upload Video / Audio",
                    file_types=[".mp4", ".mkv", ".mov", ".webm", ".avi", ".mp3", ".wav"]
                )
                btn_file = gr.Button("📂 Transcribe File", variant="primary")

        # 🌍 Language Selector
        language_selector = gr.Dropdown(
            choices=[
                "Auto Detect",
                "hi",  # Hindi (Devanagari)
                "ur",  # Urdu
                "en",  # English
                "ar",
                "fr",
                "de",
                "es",
                "ru",
                "ja",
                "zh"
            ],
            value="Auto Detect",
            label="🌍 Select Transcript Language"
        )

        output = gr.Code(label="Transcript Output", lines=15)

    btn_url.click(transcribe_media, [url_in, gr.State(None), language_selector], output)
    btn_file.click(transcribe_media, [gr.State(None), file_in, language_selector], output)

demo.launch()