Parallel Recognizer Algorithm for Automatic Speech Recognition

Research in Automatic Speech Recognition (ASR) has been very intense in recent years with focus given to accuracy and speed issues. To achieve good accuracy, the employed techniques usually rely on heavy computations. Agbago and Barri\ue8re [2] earlier defined a Three-Stage Architecture (TSA) framework for ASR composed of (1) pre-processing stage, (2) phomene recognition stage, and (3) natural language post-processor stage. Within that TSA framework, our present focus is to improve the speed of Stage 2 which looks specifically at the comparison of low level speech units. It is different from several sytems that include HMM processes in this Stage (e.g. Shawn's [5]). We present a new algorithm called Parallel Recognizer that is 320 times faster than a standard Two-Level Dynamic programming (TLDP) [3]. In comparison, working on speed at low-level, Nkagawa [4] got a reduction of 4 to 6 the time needed to compute local distances in the improved DG algorithm of Sakoe [6].La recherche en mati\ue8re de reconnaissance automatique de la parole (RAP) a \ue9t\ue9 tr\ue8s intense au cours des quelques derni\ue8res ann\ue9es, l'accent \ue9tant mis sur la pr\ue9cision et le d\ue9bit. Afin d'obtenir un bon niveau de pr\ue9cision, les techniques employ\ue9es d\ue9pendent d'habitude d'importants calculs. Agbago et Barri\ue8re [2] ont d\ue9fini plus t\uf4t un cadre d'architecture en trois \ue9tapes (TSA) pour la RAP, soit : (1) \ue9tape de pr\ue9-traitement; (2) \ue9tape de reconnaissance des phon\ue8mes; (3) \ue9tape de post-processeur en langage naturel. Dans le cadre de la TSA, nous nous concentrons ici sur la fa\ue7on d'accro\ueetre le d\ue9bit \ue0 l'\ue9tape 2, \ue9tape qui porte plus pr\ue9cis\ue9ment sur la comparaison d'unit\ue9s de parole de bas niveau. La d\ue9marche se distingue de plusieurs syst\ue8mes qui int\ue8grent des processus HMM \ue0 cette \ue9tape (p. ex. celle de Shawn [5]). Nous pr\ue9sentons un nouvel algorithme, appel\ue9 reconnaisseur en parall\ue8le, qui est 320 fois plus rapide que la programmation dynamique standard \ue0 deux niveaux (TLDP) [3]. En comparaison, pour un d\ue9bit \ue0 bas niveau, Nkagawa [4] a pu r\ue9duire de 4 \ue0 6 fois le temps n\ue9cessaire au calcul des distances locales pour l'algorithme DG am\ue9lior\ue9 de Sakoe [6].NRC publication: Ye