model_engine.py

import torch
from transformers import SpeechT5Processor, SpeechT5ForSpeechToText
import librosa

class ArabicASREngine:
    def __init__(self, model_id="MBZUAI/artst_asr_v3"):
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        print(f"Initializing ArTST Model: {model_id} on {self.device}")
        
        # Load processor and model
        self.processor = SpeechT5Processor.from_pretrained(model_id)
        self.model = SpeechT5ForSpeechToText.from_pretrained(model_id).to(self.device)

    def transcribe(self, audio_path):
        # 1. Load audio and resample to 16kHz (Requirement 4)
        speech, sr = librosa.load(audio_path, sr=16000)
        
        # 2. Pre-process features
        input_features = self.processor(
            audio=speech, 
            sampling_rate=16000, 
            return_tensors="pt"
        ).input_values.to(self.device)

        # 3. Generate IDs
        with torch.no_grad():
            predicted_ids = self.model.generate(input_features, max_length=200)
        
        # 4. Decode to Arabic Text
        transcription = self.processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
        return transcription