app.py

# app.py – Single file + player + countdown timer (ZeroGPU perfect)
import os
import gradio as gr
import spaces
from transformers import pipeline

MODEL_NAME = "palli23/whisper-small-sam_spjall"

print("Hleð Whisper módelinu...")

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

# Fix old Whisper checkpoints (required once)
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)  # 3 mínútur nóg
def transcribe_single(audio_path):
    if not audio_path:
        return None, "Hladdu upp hljóðskrá", "00:00"

    result = pipe(audio_path, chunk_length_s=30, batch_size=8)
    text = result["text"].strip()
    return audio_path, text, None  # None clears the timer when done

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

    with gr.Row():
        audio_in = gr.Audio(label="Hljóðskrá", type="filepath", waveform=True)
    
    btn = gr.Button("Transcribe", variant="primary", size="lg")
    
    with gr.Row():
        timer = gr.Timer(180, label="Tími eftir á GPU (sek)", active=True, visible=True)
    
    output = gr.Textbox(label="Útskrift", lines=20)

    # Click → transcribe + stop timer when finished
    btn.click(
        transcribe_single,
        inputs=audio_in,
        outputs=[audio_in, output, timer]
    )

demo.launch(auth=("beta", "beta2025"))