Evaluation of single-channel reverberation suppression as preprocessing for acoustic event detection
Acoustic event detection (AED) is increasingly present in applications such as health monitoring, security or home automation. A high accuracy is needed in the applications of AED to real-world scenarios. However, in an enclosed space, the source signal is corrupted by reverberation. The reverberation can be characterized by the room impulse response (RIR) and can result in a severe degradation of the AED accuracy. Therefore, dereverberation is needed to increase the robustness of AED in reverberant environment. Reverberation suppression consists in applying a real-valued gain, e.g. a Wiener gain, to the spectrogram of the input signal. The computation of this gain typically requires an estimate of the late reverberant spectral variance (LRSV). Several LRSV estimators are based on a statistical model of the RIR and the acoustical properties of the room such as the direct to reverberant ratio (DRR) or the reverberation time (T60). This paper investigates the AED performance gains by incorporating reverberation suppression using a LRSV estimator based on a statistical model of the RIR. The influence of estimation errors on the parameters input to the LRSV estimator is examined in terms of AED accuracy.