Speech Detection in Non-Stationary Noise Based on the 1/f Process

In this paper, an effective and robust active speech detectionmethod is proposed based on the 1/f process technique for signals undernon-stationary noisy environments. The Gaussian 1/f process, amathematical model for statistically self-similar random processesbased on fractals, is selected to model both the speech and thebackground noise. An optimal Bayesian two-class classifier is developedto discriminate them by their 1/f wavelet coefficients withKarhunen-Loeve-type properties. Multiple templates are trained for thespeech signal, and the parameters of the background noise can bedynamically adapted in runtime to model the variation of both thespeech and the noise. In our experiments, a 10-minute long speech withdifferent types of noises ranging from 20dB to 5dB is tested using thisnew detection method. A high performance with over 90% detectionaccuracy is achieved when average SNR is about 10dB.