can we use .wav data to do the inference directly with a pre-trained model ?