step13_arabic_ctc_transcribe.py

import os
import torch
import librosa
from transformers import AutoProcessor, AutoModelForCTC

# Arabic wav2vec2 CTC model (CPU friendly but heavy)
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-arabic"

AUDIO_PATH = "sample_trim.wav"
OUT_TXT = os.path.join("output", "asr_raw.txt")

def main():
    os.makedirs("output", exist_ok=True)

    print("Loading:", MODEL_ID)
    processor = AutoProcessor.from_pretrained(MODEL_ID)
    model = AutoModelForCTC.from_pretrained(MODEL_ID)
    model.eval()

    audio, sr = librosa.load(AUDIO_PATH, sr=16000, mono=True)
    print("Audio sec:", round(len(audio)/sr, 2))

    inputs = processor(audio, sampling_rate=sr, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(**inputs).logits

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

    # Save to file for downstream steps
    with open(OUT_TXT, "w", encoding="utf-8") as f:
        f.write(text + "\n")

    print("\n--- RAW TRANSCRIPTION ---")
    print(text)
    print(f"\nOK ✅ wrote {OUT_TXT}")

if __name__ == "__main__":
    main()