The Effects of Selected Signal Processing Techniques on the Performance of a Filter-Bank-Based Isolated Word Recognizer

To implement an isolated word recognizer based on a filter analysis, decisions must be made as to how to preprocess the speech signal prior to the filter bank analysis, how to postprocess the feature vectors obtained at the output of the filter bank analysis, and how to perform the time alignment and distance computation in the pattern comparison between an unknown test pattern and previously stored reference patterns. Often such decisions are made arbitrarily based on experience, heuristic procedures, or sometimes a few brief tests with the system. To our knowledge no one has attempted to systematically examine the effects of various signal processing techniques on the performance (as measured in word error rate) of a filter-bank-isolated word recognizer. This paper provides such a comparison by examining several of the most popular signal processing techniques and showing how they affect the performance of a particular filter bank word recognizer using telephone-quality speech.1 There are two inherent problems with any study that attempts to find the best signal processing techniques for a system via experimental means. The first is that the results presented are highly dependent on the signal processing techniques that were studied. Hence, the "optimal" way of processing the signal may not even have been investigated (due to lack of knowledge, etc). With our limited knowledge we know of no way to avoid this difficulty. The second problem is that, of necessity, each of the various signal processing techniques is studied independently of any other (thereby tacitly assuming independence of the various methods).