Sound recognition with spiking silicon cochlea and Hidden Markov Models

In this paper we explore the capabilities of a sound recognition system that combines both a novel bio-inspired custom silicon cochlea chip and a classical Hidden Markov Model (HMM). The cochlea chip front-end produces a form of representation that is analogous to the spike outputs of the biological cochlea. The system is trained with either of 2 target sounds (a clap or a bass drum) in the presence of different levels of white noise or colored noise. We provide experimental results that show 1) the system is able to detect a clap or a bass drum sound even if the amplitude of the target sound was not part of the training set and 2) the performance of the system in detecting a target sound in the presence of white noise or colored noise is around 90% for signal-to-noise ratios down to at least 0.8.