Abstract.
This paper proposes a new NN-HMM speech recognition architecture capable of real-time training. Proposed system consists of a Fuzzy ARTMAP network performing adaptive clustering of input feature vectors and a discrete HMM that models phonemes in various contexts. An internal map is used to store associations between target labels and indexes of ARTb nodes during training, thus allowing the network to generate real-life labels on recall phase. Input vectors are partitioned into three separate sets to make more flexible network vigilance and recording rate tuning. Therefore approximately equal number of ARTa clusters can be developed by each Fuzzy ARTMAP that learns a parameter set. During training phase, a supervised process of fine tuning ARTa vigilance is run to find the optimal value. This technique called Vigilance Relaxation increases global performance as "ARTa No Answer" error will gradually be minimized. Another improvement that speeds up training refers to the number of ARTa nodes constrained to be lower than a fixed limit. Only closest cluster to presented input satisfying vigilance criterion is updated when this threshold is reached. A comparison between proposed architecture and classical MLP-HMM one shows that for same global performance, Fuzzy ARTMAP labeler requires around 10% of MLP training time if the number of ARTa nodes is limited to 256.