What bad audio costs a speech model