Towards an embedded offline voice assistant using state-of-the-art hardware implementation of artificial intelligence

Deep learning is very popular because of its exceptional performance in classification. However, it is a highly complex model and its internal mechanisms are not always well understood. In addition, the parameters adjustment is often obscure . Finally, its training is long and the solutions found are not global optima. However, there is a neural network architecture that can potentially overcome some problems of deep learning: Reservoir computing. This model allows a fast training and using simple concepts of linear algebra. The solution it finds is global and deterministic. Also, it allows hardware integration, and can therefore be much more energy efficient. The aim of this thesis is to integrate an hardware neuron in the reservoir computing architecture. The device chosen to act as a neuron is a spin-torque nano-oscillator (STNO), a nano-object at the heart of spintronics research. This device offers non-linear and causal dynamics, properties conducive to neuromorphic simulation. In addition, it shares the same structure as current magnetic memory cells and is therefore compatible with CMOS technology. In order to demonstrate the lightness of such an architecture, it has been implemented on an embedded system. To show that the interaction between the software model and the hardware neuron is possible, an ASIC prototype simulating the behaviour of the STNO was implemented. The software was developed on a Raspberry Pi 4 while the ASIC was prototyped on a DE0-nano SoC including a Cyclone IV FPGA. To prove the effectiveness of the system constructed, a speech recognition task was challenged. It was a success.