app.py

# app.py — wav2vec2 multi-aug (stable + high quality)

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

import gradio as gr
import spaces
import torch
import gc
import re
import librosa

from transformers import (
    Wav2Vec2Processor,
    Wav2Vec2ForCTC
)

#MODEL_ID = "palli23/wav2vec2-icelandic-multi-aug-v2-5e-6"
MODEL_ID = "palli23/wav2vec2-icelandic-clean"
# MODEL_ID = "palli23/wav2vec2-xlsr-300m-icelandic"

# ——————————————————————————————
# Strong Icelandic cleanup
# ——————————————————————————————
def clean_text(text: str) -> str:
    text = text.lower()

    # collapse repeats (ctc artifacts)
    text = re.sub(r"(.)\1{3,}", r"\1\1", text)

    # spacing
    text = re.sub(r"\s+", " ", text)

    # punctuation spacing
    text = text.replace(" ,", ",").replace(" .", ".")
    text = text.replace(" ?", "?").replace(" !", "!")

    return text.strip()

# ——————————————————————————————
# Chunking helper (overlap improves WER)
# ——————————————————————————————
def chunk_audio(audio, sr, chunk_s=20, overlap_s=0):
    step = chunk_s - overlap_s
    chunk_len = int(chunk_s * sr)
    step_len = int(step * sr)

    for start in range(0, len(audio), step_len):
        chunk = audio[start:start + chunk_len]
        if len(chunk) < sr:  # too short
            break
        yield chunk

# ——————————————————————————————
# ZeroGPU worker
# ——————————————————————————————
@spaces.GPU(duration=180)
def transcribe_3min(audio_path):
    if not audio_path:
        return "Hlaðið upp hljóðskrá"

    processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
    model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
    model.eval().to("cuda")

    # Load audio (float32 enforced)
    audio, sr = librosa.load(audio_path, sr=16000, mono=True)
    audio = audio.astype("float32")

    texts = []

    for chunk in chunk_audio(audio, sr):
        inputs = processor(
            chunk,
            sampling_rate=16000,
            return_tensors="pt",
            padding=True
        )

        with torch.no_grad():
            logits = model(
                inputs.input_values.to("cuda")
            ).logits

        pred_ids = torch.argmax(logits, dim=-1)
        text = processor.batch_decode(pred_ids)[0]
        texts.append(text)

    final_text = clean_text(" ".join(texts))

    # Cleanup (critical)
    del model
    del processor
    gc.collect()
    torch.cuda.empty_cache()

    return final_text

# ——————————————————————————————
# UI
# ——————————————————————————————
with gr.Blocks() as demo:
    gr.Markdown("# Íslenskt ASR – wav2vec2 (multi-aug)")
    gr.Markdown("**stöðugt · chunked · post-processed**")
    gr.Markdown("**Hafa samband:** pallinr1@protonmail.com")

    audio_in = gr.Audio(type="filepath", label="Hlaðið upp .mp3 / .wav")
    btn = gr.Button("Transcribe", variant="primary", size="lg")
    output = gr.Textbox(lines=20, label="Útskrift")

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

demo.launch(
    share=True,
    server_name="0.0.0.0",
    server_port=7860,
)