app.py

import gradio as gr
from transformers import pipeline
import torch

# Detect device and dtype for efficiency/memory
device = "cuda:0" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if torch.cuda.is_available() else torch.float32

# Load all 4 models (with chunking for long audio)
pipe1 = pipeline(
    "automatic-speech-recognition",
    model="IJyad/whisper-large-v3-Tarteel",
    torch_dtype=dtype,
    device=device,
    chunk_length_s=30,
)
pipe2 = pipeline(
    "automatic-speech-recognition",
    model="deepdml/whisper-medium-ar-quran-mix-norm",
    torch_dtype=dtype,
    device=device,
    chunk_length_s=30,
)
pipe3 = pipeline(
    "automatic-speech-recognition",
    model="naazimsnh02/whisper-large-v3-turbo-ar-quran",
    torch_dtype=dtype,
    device=device,
    chunk_length_s=30,
)
pipe4 = pipeline(
    "automatic-speech-recognition",
    model="Habib-HF/tarbiyah-ai-whisper-medium-merged",
    torch_dtype=dtype,
    device=device,
    chunk_length_s=30,
)

def transcribe(audio):
    if audio is None:
        return "No audio", "No audio", "No audio", "No audio"
    
    # Force Arabic language for consistency (these models are Arabic/Quran specialized)
    kwargs = {"language": "arabic", "task": "transcribe"}
    
    text1 = pipe1(audio, generate_kwargs=kwargs)["text"]
    text2 = pipe2(audio, generate_kwargs=kwargs)["text"]
    text3 = pipe3(audio, generate_kwargs=kwargs)["text"]
    text4 = pipe4(audio, generate_kwargs=kwargs)["text"]
    
    return text1, text2, text3, text4

with gr.Blocks(title="Quran Whisper Models Comparison") as demo:
    gr.Markdown("""
    # Quran ASR Models Comparison
    
    Upload or record a short Quranic recitation and compare transcriptions side-by-side.
    
    Models:
    - IJyad/whisper-large-v3-Tarteel (large-v3, high accuracy)
    - deepdml/whisper-medium-ar-quran-mix-norm (medium)
    - naazimsnh02/whisper-large-v3-turbo-ar-quran (turbo, fast & accurate)
    - Habib-HF/tarbiyah-ai-whisper-medium-merged (medium, merged general + Quran)
    """)
    
    audio_input = gr.Audio(
        sources=["microphone", "upload"],
        type="filepath",
        label="Record from mic or upload audio file (WAV/MP3, preferably Quran recitation)"
    )
    
    btn = gr.Button("Transcribe with all 4 models")
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("### IJyad/whisper-large-v3-Tarteel")
            out1 = gr.Textbox(label="Transcription", lines=6, rtl=True)
        with gr.Column():
            gr.Markdown("### deepdml/whisper-medium-ar-quran-mix-norm")
            out2 = gr.Textbox(label="Transcription", lines=6, rtl=True)
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("### naazimsnh02/whisper-large-v3-turbo-ar-quran")
            out3 = gr.Textbox(label="Transcription", lines=6, rtl=True)
        with gr.Column():
            gr.Markdown("### Habib-HF/tarbiyah-ai-whisper-medium-merged")
            out4 = gr.Textbox(label="Transcription", lines=6, rtl=True)
    
    btn.click(transcribe, inputs=audio_input, outputs=[out1, out2, out3, out4])
    
    gr.Markdown("""
    **Notes:**
    - Best for short Quran recitations (mic recordings are usually <30s).
    - Transcriptions are plain Arabic text (no tashkeel/diacritics in most cases).
    - GPU highly recommended — CPU will be slow.
    - These models are Quran-specialized; general Arabic speech may not work well.
    """)

demo.queue()  # Helps with concurrent users
demo.launch()