import gradio as gr
import torch
import librosa
import numpy as np
from transformers import AutoProcessor, AutoModelForCTC

# Load model and processor
print("Loading model...")
processor = AutoProcessor.from_pretrained("HAMMALE/mms-darija-finetuned")
model = AutoModelForCTC.from_pretrained("HAMMALE/mms-darija-finetuned")

def transcribe_audio(audio_file):
    try:
        # Load audio
        if audio_file is None:
            return "Please upload an audio file."
        
        # Load and preprocess audio
        audio, sr = librosa.load(audio_file, sr=16000)
        
        # Handle very short audio
        if len(audio) < 1600:  # Less than 0.1 seconds
            return "Audio too short. Please upload a longer audio file."
        
        # Process with model
        inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
        
        # Inference
        with torch.no_grad():
            logits = model(**inputs).logits
        
        predicted_ids = torch.argmax(logits, dim=-1)
        transcription = processor.batch_decode(predicted_ids)[0]
        
        return transcription if transcription.strip() else "No transcription generated."
        
    except Exception as e:
        return f"Error processing audio: {str(e)}"

# Create Gradio interface
demo = gr.Interface(
    fn=transcribe_audio,
    inputs=gr.Audio(type="filepath", label="Upload Darija Audio"),
    outputs=gr.Textbox(label="Transcription", placeholder="Transcription will appear here..."),
    title="🎤 Darija Speech Recognition",
    description="Upload an audio file in Moroccan Arabic (Darija) and get the transcription. This model was fine-tuned on the Darija Bible dataset.",
    article="Model: [HAMMALE/mms-darija-finetuned](https://huggingface.co/HAMMALE/mms-darija-finetuned)",
    examples=[
        # You can add example audio files here if you have them
    ],
    cache_examples=False,
    theme=gr.themes.Soft()
)

if __name__ == "__main__":
    demo.launch()