app.py

# app.py – VIRKAR Á ÖLLUM Spaces (jafnvel gömlum Gradio)
import os
import gradio as gr
import spaces
from transformers import pipeline

MODEL_NAME = "palli23/whisper-small-sam_spjall"

print("Hleð Whisper módelinu einu sinni...")

pipe = pipeline(
    "automatic-speech-recognition",
    model=MODEL_NAME,
    torch_dtype="auto",
    device="cuda",
    token=os.getenv("HF_TOKEN")
)

# Fix fyrir gamlar Whisper útgáfur
if not hasattr(pipe.model.generation_config, "lang_to_id") or pipe.model.generation_config.lang_to_id is None:
    pipe.model.generation_config.lang_to_id = {"is": 50259}
    pipe.model.generation_config.task_to_id = {"transcribe": 50359, "translate": 50358}
    pipe.model.generation_config.forced_decoder_ids = None

print("Módel tilbúið!")

@spaces.GPU(duration=180)
def transcribe_single(audio_path):
    if not audio_path:
        return None, "Hladdu upp hljóðskrá fyrst", "00:00"

    result = pipe(audio_path, chunk_length_s=30, batch_size=8)
    text = result["text"].strip()
    return audio_path, text, None  # Slekkur á timer þegar búið

with gr.Blocks() as demo:
    gr.Markdown("# Íslenskt Whisper – Mjög lágt WER")
    gr.Markdown("Hladdu upp einni skrá (allt að 5 mín) → Transcribe")

    audio_in = gr.Audio(label="Hljóðskrá", type="filepath")  # Virkar á öllum Gradio útgáfum
    btn       = gr.Button("Transcribe", variant="primary", size="lg")

    # Einfaldur timer án label/active/visible (virkar á Gradio 3.x)
    timer     = gr.Timer(value=180)  

    output    = gr.Textbox(label="Útskrift", lines=20)

    btn.click(
        transcribe_single,
        inputs=audio_in,
        outputs=[audio_in, output, timer]
    )

# Login: beta / beta2025
demo.launch(auth=("beta", "beta2025"))