In this work, we propose a frame selection scheme based on the smoothed instantaneous energy of samples, local order statistics for them, and average of a binary energy indicator over the frame to measure the reliability of frames. By selection of reliable frames, a four-stage feature normalization and transformation process is further proposed: mean normalization, variance normalization, first-stage principal component analysis, and multi-eigenvector temporal filtering. Extensive experiments verified that the performance of each individual stage can be significantly improved by the proposed frame selection scheme, and the overall performance can also be improved stage by stage for all types of noise and all SNR values defined in AURORA 2.