Abstract:
We describe ``WarpNet'', a time-warping algorithm for speech recognition based on neural nets. WarpNet is a self-organizing matching mechanism intended to replace dynamic programming (Dynamic Time Warping or Viterbi-algorithms). The concept is based on elastic models that self-organize in the spirit of Self Organizing Maps and Optimizing Maps \cite{Peterson96} so as to minimize the the sum of distances over two utterances, or any time series. We show that both the recognition accuracy, and the computing time of WarpNet are comparable to the optimal matching method, dynamic programming.