Neural network device for recognizing speech commands based on hardware accelerators of the NEUROMATRIX family

Abstract

Various approaches to solving the problem of speech command recognition in real time using artificial neural networks (ANN) of several types are considered. A brief overview of deep fully-connected, convolutional and recurrent neural networks is given. It is shown that ANNs of convolutional and recurrent types are capable of providing the required level of performance and accuracy when performing inference in real time. A comparison is given of the performance of convolutional and recurrent algorithms when ported to the K1879VM6Ya hardware platform from the NeuroMatrix family of high-performance digital signal processors. When porting, the features of the microprocessor architecture were considered. All basic operations are performed on a vector matrix coprocessor. The advantage of convolutional ANN has been demonstrated. Considering the results of research into a neural network of this type on the K1879VM6Ya platform, it was transferred to a more productive hardware accelerator - the K1879VM8Ya system-on-chip. The prospects for creating a speech recognition system on the K1879VM8Ya platform are considered.

Keywords

• speech recognition, artificial neural network, convolutional neural network, microprocessor, k1879vm8ya, k1879vm6ya.