Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
In this paper we present an Unconstrained Level Building (ULB) algorithm based on the known Level-Building approach , which allows to deal, in a new particular form, the problem of the interior pauses or silences produced when strings of digits are pronounced in a natural way (strings of digits may contain silences or pauses between digits, due, for example, to the speaker's need of breathing, or to his hesitation in recalling or reading the next digit to pronounce). To evaluate the performance of the new algorithm a speaker dependent connected digits Spanish data base has been used, consisting of 129 digit strings, with lengths ranging from one to five digits, pronounced by a male speaker. The recognition set contains 46 strings, and the training set 64. The 19 remaining strings have been used, as an independent testing set, to control the convergence of the training procedure. Several experiments were carried out to determine the best set of conditions for recognition of digit strings. The improvement obtained on String Accuracy using the ULB algorithm instead of a Level Building with an HMM for the silence is 4.3%, and without any silence modelling is nearly 11%.
Bibliographic reference. Cifuentes, S. / Colas, J. / Savoji, M. / Pardo, Josť M. (1992): "A new algorithm for connected digit recognition", In ICSLP-1992, 1499-1502.