Título: Rule-based stop recognition with the Karhunen-Loeve transform
Autores: Morgan, Siân Yvette
Fecha: 1994
Publicador: McGill University - MCGILL
Fuente:
Tipo: Electronic Thesis or Dissertation
Tema: Engineering, Electronics and Electrical.
Descripción: This thesis proposes a speaker-independent, continuous-speech stop recognition algorithm. The method incorporates knowledge into the system by means of rules whose conditions are linked to acoustic events. Coarticulatory effects present in the speech are taken into account with the analysis of the transient portions of the signal.
The novelty of this approach consists of the use of the Karhunen-Loeve transform to segment the speech and to extract both steady-state and dynamic features. Sequences of stops followed by vowels were segmented with the second Karhunen-Loeve transform coefficient, whereas information about the formant transitions and energy distribution was extracted by means of the basis vectors of the transformed space.
Tests of the proposed rule-based system resulted in 71% correct classification of the stops' place of articulation, and 80% correct classification of the stops' manner of voicing.
More complex systems have previously attained better results. However, the simplicity and computational cost efficiency of this approach would be of considerable benefit to any large ASR system.
Idioma: en