Neural networks have been shown to be a powerful approach to speech recognition. Especially the class of time-delay neural networks (TDNN) exhibiting a time invariance feature has been a popular method. In this paper we emphasize another approach to speech recognition and analysis where neural nets are used differently from principles like TDNN in the way in which time is treated. We call this approach time-event neural networks (TENN) because it is based on the detection of events as time moments or intervals of interest as the first step and recognition (classification) of the events as the second step. These steps may also be integrated together. The advantages are explicit temporal information gained from events and reduced computation due to focused classification. The paper describes primarily the general principles of the approach. Experiments and preliminary results are shown from diphone-based speech recognition. Keywords: Speech recognition; Neural networks; Event-Based Speech Analysis
Bibliographic reference. Aliosaar, Toomas / Karjalainen, Matti (1991): "Event-based recognition and analysis of speech by neural networks", In EUROSPEECH-1991, 1031-1034.