Low-complexity in-sensor audio detection with experimental validation

In this paper, we present a low-complexity detection algorithm for the recognition of different audio signal patterns. The proposed detection algorithm evolves through two main processing phases: (a) coarse and (b) fine. The evolution between these two phases is described through a finite state machine (FSM) model. The use of different processing phases is expedient to reduce the computational complexity, thus making our algorithm suitable for wireless sensor networking scenarios, where the in-sensor energy consumption needs to be kept as low as possible. In fact, fine processing (in the frequency domain) is carried out only when an “atypical” audio signal is detected. On the other hand, coarse processing (in the time domain), performed a larger number of times, has a much lower complexity. The proposed approach is validated through audio signals experimentally acquired with a commercial microphone, embedded in a wireless sensor node. The obtained results show that our processing technique allows to detect efficiently the presence of signals of interest (identified by properly selected spectral signatures) and to reliably distinguish different audio signal patterns, e.g., speech and non-speech audio signals.