Exploring the Back-Propagation Network for Speech Applications

Abstract : The goal of our research is to explore how back-propagation networks, trained to learn the significant representations of preprocessed speech, affect novel speech data. Using networks of different sizes with different preprocessing methods, we hope to discover the features learned and how this information may aid the performance of difficult speech processing tasks. Neural networks have sophisticated abilities for processing and filtering signals. In particular, Elman and Zipser demonstrated that the back-propagation network develops significant feature representations which may be useful for both segmenting and recognizing speech. Such networks might find applications in speech compression and/or speech normalization. The network's apparent potential for speech applications justifies further exploration, and this paper describes our work in process.