Improvements in isolated word recognition

For isolated word recognition, possibilities for improving the recognition accuracy are investigated for a given feature extraction, which is based on a short term spectrum analysis by means of band-pass filtering. A number of preprocessing steps are discussed, which are to be applied prior to time alignment via dynamic programming. These preprocessing steps include normalization of short term spectra with respect to the long term spectrum, amplitude normalization and spectral channel contour smoothing. For nonlinear time alignment, a method based on spectral change is investigated, alone and in combination with dynamic programming. The resulting distance measure is incorporated into pattern recognition schemes according to the minimum distance and nearest neighbor principles. The different processing steps are evaluated in a speaker dependent mode of operation separately for two vocabularies: the ten German digits and twelve major German airport city names. In comparison with the use of standard mean normalization of short term spectra and dynamic programming, the aforementioned techniques allow for a performance improvement in terms of error rate reduction by a factor of 3-5, while at the same time offering savings in computing time and reference memory requirements by a factor of 10 and 3, respectively.