A new network model, U-net, is proposed to recognize continuous speech, based on the non-uniform unit which is a kind of acoustic sub-word unit. In this model, input speech can be segmented into units by using a part of the network before classification. The unit has steady states at the boundaries and a transient state in the middle. The network structure is designed according to the structure of the unit. The steady states and transient state are recognized by separate networks and different feature parameters are used. For the transient part a delta parameter is used. The segmentation net is trained to reduce the number of unit classes.
Bibliographic reference. Yu, Ha-Jin / Oh, Yung-Hwan (1995): "A neural network using non-uniform units for continuous speech recognition", In EUROSPEECH-1995, 1677-1680.