Explore the quality of reconstruction when audio is tokenized using various number of code books in the Mimi model.