Many-Speakers Single Channel Speech Separation with Optimal Permutation Training