Recognition of speaker-independent isolated Persian digits using an enhanced vector quantization algorithm

Vector quantization (VQ) is a fast and simple classification algorithm that has been widely employed for the recognition of isolated spoken words. However, this algorithm and most of its improved versions fail to accurately distinguish words with similar vowels. The spoken pattern of digits/dow/ and/noh/ (2 and 9 respectively) in Persian is a good example for this type of similarity. In this paper we have proposed an enhanced vector quantization algorithm in which the deterministic role of the short consonants at the beginning of the words is taken into account. In this algorithm an unknown vector is judged based on the classification results of two set of codebooks. The first set of codebooks is constructed by the initial portions of the words while the other set is constructed by the whole words. The performance of the proposed algorithm was experimentally verified against other VQ-based algorithms. While the overall performance of the proposed algorithm was above the others, in the case of similar words it could remarkably decrease the number of misclassification. This improvement was achieved by only a small increase in the computational load.