On the use of high order derivatives for high performance alphabet recognition

In this paper I propose new feature vectors for automatic speech recognition. They are based on Mel-cepstrum vectors augmented by derivatives. In the literature, many systems using just two derivatives—delta and delta delta—are described. But none explores the use of higher order derivatives. This paper presents alphabet recognition results on the Isolet database, using feature vectors containing up to the fifth-order derivatives. For this paper I did not use the HTK toolkit proposed by Cambridge University. I developed my own HMM system. I show that with vectors incorporating all the derivatives up to the fifth one, 97.54% mean recognition accuracy was achieved, result which is comparable to the best published one on this database (97.6%), if the recognition accuracy confidence interval concerning this task (approximately 0.3%) is taken into account. It is important to note that this result was obtained without segmenting the speech files by an endpoint detection algorithm. This is an unfavourable experimental condition compared to previous published research works. As a consequence, my system is one of the most powerful systems ever implemented for alphabet recognition.