Hybrid neuromorphic system for automatic speech recognition

A multilayer neural network, equipped with a two‐memristors synapse, for speech recognition is proposed. The discussed neuromorphic neural network is a hybrid system which uses a Gaussian–Bernoulli restricted Boltzmann Machine (RBM) to transform the speech data into sparse encoded binary data. The sparse data is used to train a standard RBM, and a two‐memristors synapse using AI/Pr 0.7 Ca 0.3 MnO 3 memristor is used as a connection between the two layers of the RBM. The simulations are performed with the real memristor's behavioural data, for potentiation and depression, to adjust learn‐able parameters of the neuromorphic RBM. Instead of hard coded representation of the data, memristive synapses follow the biological way of learning for plasticity by training with examples. Experimentation with the American common use vowels and its results assert the efficacy of the proposed architecture.