The ASPEN Parallel Computer, Speech Recognition, Parallel Dynamic Programming

In previous work [1], the DADO/DSP multiprocessor architecture and parallel dynamic programming (PDP) were proposed as an approach to accelerating large-vocabulary connected-speech recognition. In this work, we describe the important characteristics of the DADO/DSP that make it well-suited for miniaturization, review the principles of applying PDP to speech recognition, and report on the results of experiments executing PDP that support these principles. The problem of developing a multiprocessor architecture and parallel algorithms for large vocabulary speech recognition is important for two reasons. The first is that it is a computationally intensive problem that can require orders of magnitude acceleration over uniprocessors to achieve real-time performance. The second is that there is still much algorithm development work to be done, which requires a programmable computer rather than a fixed implementation.