The first part of this paper describes an algorithm using dynamic programming which allows endpoint relaxation and then achieves an implicit segmentation of the pattern to be compared.

In the second part, we introduce vector quantization in order to reduce the memory size occupied by data (one or several patterns for each word of the vocabulary). We propose several recognition methods using dynamic time warping and we compare their performances.

In the last part, we extend the Bridle and Nakagawa algorithm by using endpoint relaxation, syntactic constraints, vector quantization and we propose a method which takes into account liaisons and coarticulation effects at the boundaries of connected words.