Review of TDNN (time delay neural network) architectures for speech recognition

The TDNN architecture for speech recognition is described, and its recognition performance for Japanese phonemes and phrases is explained. In comparative studies, it is shown that the TDNN yields superior phoneme recognition performance. The TDNN optimized for phoneme recognition, however, does not necessarily result in optimized word or phrase recognition performance, as overfitting to the specific phoneme data or recording conditions may occur. Care must therefore be taken to achieve robust integration, and several studies toward this goal are reported