Psychoacoustic models are very good, but MP3 has unavoidable coding weaknesses. Use a modern audio codec and you won't have this problem nearly as much.