It is well know that the pattern matching approach to word-based recognition yields good results when the time aligriment is carried out by means of dynamic programming. However, the high computational expenditure needed by it, specially for large vocabularies, has motivated a number of works proposing the use of techniques that reduce the pattern lenght. In this paper, three different frame or spectrum reduction methods are investigated in a LPC isolated word recognition framework. The reduction performance with different degrees of nonlinearity is tested, and reasons showing its relationship with some kind of actual endpoint detection innaruracies are given.

Peer Reviewed