app.py

# app.py — Clean output, no timestamps crash, stronger anti-hallucination

import os
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"

import gradio as gr
import spaces
from transformers import pipeline
import torch
import gc
import re

# Post-processing to remove common ASR junk
def clean_asr_text(text: str) -> str:
    if not text:
        return ""
    
    # Remove bracketed/angled garbage
    text = re.sub(r'<[^>]+>', '', text)                     # <UNK>, <|0.00|>, etc.
    text = re.sub(r'\[.*?\]', '', text)                     # [HIK:xxx], [laughter], etc.
    
    # Extra Icelandic/common noise patterns
    text = re.sub(r'(?i)\b(unk|hik|laughter|music|cough|applause|noise|background)\b', '', text)
    
    # Normalize spaces & punctuation
    text = re.sub(r'\s+', ' ', text)
    text = re.sub(r'^\s+|\s+$', '', text)
    text = text.replace(' ,', ',').replace(' .', '.').replace(' ?', '?').replace(' !', '!')
    text = re.sub(r' +([.,!?])', r'\1', text)
    
    return text.strip()

# ──────────────────────────────────────────────
# Transcription function
# ──────────────────────────────────────────────
@spaces.GPU(duration=180)
def transcribe_3min(audio_path):
    if not audio_path:
        return "Hlaðið upp hljóðskrá"

    pipe = pipeline(
        "automatic-speech-recognition",
        model="palli23/whisper-tiny-distilled-spjallromur-polish-v5",
        torch_dtype=torch.float16,
        device=0,
    )

    try:
        result = pipe(
            audio_path,
            chunk_length_s=30,
            batch_size=8,
            return_timestamps=False,           # ← avoids generation_config crash
            generate_kwargs={
                "num_beams": 5,
                "repetition_penalty": 1.3,     # stronger than before
                "no_repeat_ngram_size": 4,     # prevents repeating phrases
                "temperature": 0.0,            # greedy decoding = least creative garbage
                "suppress_tokens": [-1],       # try to block <unk>
                "max_new_tokens": 444,         # safety limit
            }
        )

        raw_text = result.get("text", "")
        cleaned = clean_asr_text(raw_text)

    except Exception as e:
        cleaned = f"Villa við umritun: {str(e)}"

    # Clean up GPU memory
    del pipe
    gc.collect()
    torch.cuda.empty_cache()

    return cleaned or "(ekkert texti fannst eða villa kom upp)"

# ──────────────────────────────────────────────
# Gradio UI
# ──────────────────────────────────────────────
with gr.Blocks() as demo:
    gr.Markdown("# Íslenskt ASR – 3 mínútur (hreinsuð útgáfa)")
    gr.Markdown(
        "**Model:** palli23/whisper-tiny-distilled-spjallromur-polish-v5  \n"
        "**Stillingar:** no timestamps, temperature=0.0, repetition_penalty=1.3, no_repeat_ngram_size=4  \n"
        "Reynir að fjarlægja <UNK>, [HIK...], [laughter] o.s.frv."
    )
    gr.Markdown("**Hafa samband:** pallinr1@protonmail.com")

    audio_in = gr.Audio(
        type="filepath",
        label="Hlaðið upp .mp3 / .wav / .m4a (allt að ~5 mín)",
        format="mp3"
    )
    
    btn = gr.Button("Umrita", variant="primary", size="lg")
    
    output = gr.Textbox(
        lines=25,
        label="Útskrift (hreinsuð)",
        placeholder="Hér kemur textinn..."
    )

    examples = gr.Examples(
        examples=[
            ["example_clip_14nov2025.mp3"],
        ],
        inputs=audio_in,
        label="Dæmi (ef þú hefur sett upp dæmi skrá)"
    )

    btn.click(
        fn=transcribe_3min,
        inputs=audio_in,
        outputs=output
    )

# ──────────────────────────────────────────────
# Launch
# ──────────────────────────────────────────────
if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,               # ← set to True locally if you want public link
        debug=False
    )