Skip to content. | Skip to navigation

Personal tools
Log in

Forgot your password?
You are here: Home Publications Abstracts Spike-based sound recognition using neuromorphic VLSI hardware

Spike-based sound recognition using neuromorphic VLSI hardware

by mabdollahi last modified Sep 26, 2011 12:20 PM
Abstract of the poster presented at the "Champalimaud Neuroscience Symposium", Lisbone, Champalimaud Center for the Unknown, September 18-21 2011.

Symposium website: Champalimaud Neuroscience Symposium.


Spike-based sound recognition using neuromorphic VLSI hardware


Mohammad Abdollahi, Fabio Stefanini, Shih-Chii Liu, Hynek Hermansky, Giacomo Indiveri


We are implementing spike-based models of sound recognition using neuromorphic VLSI technology. Our goal is to develop a hardware sound recognition system which is based on continuous time, stimulus-driven, asynchronous, distributed collective and self-organizing principles. The front-end of this system is implemented using an event-based silicon cochlea, while the output stage will be implemented using a multi-neuron chip with spike-based learning properties. The processing of auditory signals at the intermediate computational stages will be implemented using state-of-the-art spiking neural models, based on Spike-Timing Dependent Plasticity (STDP) learning methods and exploiting the dynamics of recurrently connected networks of neurons. As the first step and in order to compare the silicon cochlea as a speech recognition front-end to the state-of-the-art alternatives, and to evaluate the discriminative information present in the signals extracted by silicon cochlea we carried out speaker-independent isolated digit recognition experiments. The results show promising recognition accuracies (over 95%) across a large number of different speakers from TIDIGITS database, despite all the artifacts present in the hardware implementation and its very limited input dynamic range: the discriminative information in spike patterns is sufficient for a task as complex as speaker-independent isolated keyword recognition.