3 VERIFICATION
The goal of the verification is to exploit the strengths of the HMMs to overcome the shortcomings in order to improve the recognition rate of the HRS under the constraint of not increasing the recognition time. The approach we have selected for this is to develop the concept of segmental neural networks as proposed in speech recognition by Zavaliagkos et al. [30]. A segmental neural network is a neural network that estimates a posteriori probability for a complete character segment as a single unit, rather than a sequence of conditionally independent observations. Furthermore, the verification approach was chosen to be modular and applied only to the output of the recognition system. Such a modular approach is suitable for further evaluating the effects of the verification on the overall recognition performance.