DEEP BOTTLENECK FEATURES AND SOUND-DEPENDENT I-VECTORS FOR SIMULTANEOUS RECOGNITION OF SPEECH AND ENVIRONMENTAL SOUNDS