Improved parameter tying for efficient acoustic model evaluation in large vocabulary continuous speech recognition.