Tunable time delay neural networks for isolated word recognition

Describes a new neural network structure and a corresponding new sequential training technique for speech recognition. The proposed system is a modification of the original time delay neural network (TDNN) structure of Waibel et al. [1989]. The new structure consists of a group of sub-nets, and each isolated word or phoneme to be recognized corresponds to one sub-net. Since each sub-net deals with only one recognition unit, it may be trained independently. Each sub-net is a TDNN which the authors train with a new sequential training algorithm. The system has attained close to 100% accuracy for a multi-speaker, isolated word recognition task and 86.44% accuracy for a three voiced-stop-consonants ("B", "D" and "G"), speaker-independent phoneme recognition task. Results for phoneme recognition compared favorably with the best result obtained by Bryant [1992] using Sawai's block windowed neural network architecture with improvement by 14.44% for the same task.