A neural network architecture for speech segmentation using mean field annealing

As a dual algorithm to the Geiger-Girosi restoration scheme, a new segmentation method is introduced and used to demonstrate an approach to phoneme boundary detection. Also the authors introduce a neural network suitable for this algorithm, which consists of sigmoid neurons and Sigma-Pi neurons. Experimental results show that the new algorithm is superior to the forward-backward algorithm and the Geiger-Girosi algorithm in terms of position accuracy and recognition accuracy as well as computational speed for phoneme-boundary detection.<
>