A unified deep modeling approach to simultaneous speech dereverberation and recognition for the reverb challenge