Нейросетевой распознаватель фонем русской речи на мультипроцессорной графической плате

Abstract

In the paper the usage of the modern graphics card sw/hw tools with the parallel architecture for construction of the Russian speech neural network phoneme recognizer is considered. Decomposition of neural network algorithms into fragments for its mapping on the multiprocessor system is offered. The carried out experiments showed advantage of the parallelization neurocomputing on a graphics card for automatic speech recognition in real time.