Exploiting speech knowledge in neural nets for recognition

This paper argues that neural networks are good vehicles for automatic speech recognition not simply because they provide non-linear pattern recognition but because their architecture allows the incorporation and exploitation of existing knowledge about speech. The paper is in two parts: Part I defends the need for the incorporation of existing knowledge, while Part II sketches a speech recognition architecture that uses neural networks to represent and exploit existing phonological and linguistic knowledge. The first part of the paper argues that the definition of the speech recognition problem implies that prior knowledge of linguistic analysis is essential for its solution, and suggests that the currently poor exploitation of such knowledge is a consequence of contemporary pattern recognition architectures. Criticism is made of the current emphasis on syntactic pattern recognition algorithms operating at the level of the phonetic segment. The second part of the paper demonstrates that a network architecture for the lexicon provides a mechanism for the incorporation and exploitation of a range of phonological analyses. Furthermore, through the explicit separation of phonological representations from phonetic ones, there exists the possibility of constructing a front-end phonetic component on purely pattern recognition principles. Through normalisation of speaker and environment, the phonetic component may be interfaced to the network lexicon to provide a complete recognition architecture which avoids compromise in the exploitation of speech knowledge.