app.py

import gradio as gr
import librosa
import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# Load pre-trained model and processor directly from Hugging Face Hub
model = Wav2Vec2ForCTC.from_pretrained("boumehdi/wav2vec2-large-xlsr-moroccan-darija")
processor = Wav2Vec2Processor.from_pretrained("boumehdi/wav2vec2-large-xlsr-moroccan-darija")

def transcribe_audio(audio):
    # Load the audio file from Gradio interface
    audio_array, sr = librosa.load(audio, sr=16000)
    
    # Tokenize the audio file
    input_values = processor(audio_array, return_tensors="pt", padding=True).input_values
    
    # Get the model's logits (predicted token scores)
    logits = model(input_values).logits
    
    # Get the predicted tokens
    tokens = torch.argmax(logits, axis=-1)
    
    # Decode the tokens into text
    transcription = processor.decode(tokens[0])
    
    return transcription

# Create a Gradio interface for uploading audio or recording from the browser
demo = gr.Interface(fn=transcribe_audio, 
                    inputs=gr.Audio(source="upload", type="filepath"), 
                    outputs="text")

demo.launch()