Optimization Scheme of Blind Signal Extraction Based on Joint Noise Suppression and Dereverberation
Fine Dwinita Aprilyanti (1151127)
In this presentation, I address an optimization problem for blind signal extraction with joint noise suppression and dereverberation. Hands-free speech interface provides convenient and natural human-machine interaction. On the other hand, as the distance between speaker and microphones array increases, the system suffers from adverse effect of background noise and reverberation. Recently, a method to jointly suppress background noise and late reverberation has been proposed, combining frequency domain blind signal extraction (FD-BSE) and Wiener filter (WF). However, some parameter values in this method are still chosen manually, which make this method still impractical to be implemented in real environment.
In this study, I proposed an optimization scheme of joint noise suppression and dereverberation. The proposed scheme utilizes the amount of generated musical noise as control parameter, which is assessed by higher-order statistics measures, namely, noise kurtosis ratio. To maintain the optimum performance for automatic speech recognition (ASR) implementation, I also proposed an optimization scheme which utilizes the maximum likelihood of acoustic model as additional parameter. The effectiveness of the proposed schemes is evaluated through experimental evaluation and recognition task results.