A speaker-independent isolated-word recognizer based on polynomial classifiers

The author is concerned with speaker-independent isolated-word recognition for ordering over telephone using mean square polynomial classifiers (MSPC). This classification method has proved to be very efficient for pattern recognition and other classification tasks; therefore the author investigated the MSPC for our recognition task. He compares different MSPC structures with each other and with a widely used method, multiple-template dynamic time warping (DTW). Based on these results, the author implements the recognizer for the desired classification task as a two-stage hierarchical MSPC. The recognizer has to accommodate not only a wide population of speakers, but also various transmission conditions. To adapt the recognizer to various microphones and transmission lines, he filters a 108-speaker training sample systematically with a representative set of simulated frequency responses in order to increase the effective number of training samples. The resulting classifier was found to be insensitive to a wide range of frequency conditions