Speech processing and recognition using integrated neurocomputing techniques (Esprit Basic Research Action 3228: SPRINT)

Abstract

This paper is a brief summary of the first deliverable of SPRINT project, partially supported by ESPRIT Basic Research Action (ESPRIT BRA - 3228). The project, launched on April 1989, aimed at tackling various problems that remain unsolved in speech recognition by exploring the particularities of neural networks (e.g. nonlinearity, self-organization, parallelism) to extend the capability profile of automatic speech recognition systems in important directions, in terms of adaption to new environments (e.g. new speakers, channels), and noise reduction. Speech is a complex phenomenon but it is useful to divide it into levels of representation. In this framework the performance of neural networks at each level will be evaluated with the aim to solve pr3blems that occur at that specific level while bearing in mind the general problem of improving the recognizer capability profile. For that purpose connectionism paradigms and particularities are exploited to tackle the major problems in relationship with speech variabilities: new speakers and/or new environments. The approach adopted should allow us to better understand the behaviour of neural networks when applied to speech, to appreciate their usefulness, and in the future to efficiently implement and use them in speech recognition devices in order to tackle speech variabilities (inter-speakers, noise). The work described below concerns speaker adaptation, noise immunity, classification of speech parameters using a set of “phonetic” symbols, investigation of relevant graphemic symbols, and classification of a sequence of feature vectors by lexical access.