The Tdrbf: a Shift Invariant Radial Basis Function Network

| Conventional speech recognition systems based on Multi Layer Percep-trons often use Time Delay Neural Networks (TDNN). TDNNs were rst used for speech recognition by Waibel et al., but long training times and large numbers of parameters that need careful adjustment make it hard to achieve good performance. In contrast, networks using Radial Basis Functions (RBF) can be constructed systematically and training is signiicantly faster than Back Propagation for TDNNs. However, RBF networks are neither shift invariant in time, nor can they detect features in time. A Time Delay Radial Basis Function Network (TDRBF) for shift invariant recognition of features in time is presented in this paper. The TDRBF combines characteristics of the Time Delay Neural Networks and Radial Basis Functions. The ability to detect features and their temporal relationship independent of position in time is inherited from TDNN. The use of RBFs leads to shorter training times and fewer parameters to adjust for best performance, which makes it much easier to apply TDRBF to new tasks. In addition, the underlying training algorithm chooses the required number of prototypes automatically. In a task of recognizing three diierent pho-nemes, it is shown that while the performance of TDRBF and TDNN are comparable , the TDRBF network is signiicantly faster to train.