摘要：The choice of the best feature set remains the main issue for the successful speech recognition system. In the literature quality of features is estimated by calculating the classification error. So that, it is needed to run classification process with each explored feature system in order to choose the highest quality one. Therefore, a major issue of this paper is to propose the methodology for speech feature quality establishment without running the classification process. The proposed methodology is based on metrics that do not need parameters setting, thus the results can be uniformly interpreted across the different problems. The methodology consists of two parts: 1) establishment of the best metric in combination with used classifier, 2) making decision regarding the highest quality feature system. In the experiment we use Dynamic Time Warping (DTW) classifier. The metric of intra/inter class nearest neighbor distances (Q3) is identified as the best metric. Employing our proposed methodology we established Perceptual Linear Prediction analyses to be the highest quality feature system within the explored feature systems set. The correctness of the results is confirmed by DTW classification error.
