Robust speech separation using two-stage independent component analysis
Parham Aarabi, Sam Mavandadi · 2003
This paper proposes a two stage speech (8, 21. Standard ICA, however, suffers from several lim- separation architecture involving a microphone selec- itations. First, there is assumed to be as many ideal tion stage and an Independent Component Analysis mixed signals available as there are independent sig- (ICA) stage. Standalone ICA often fails to correctly nals. In practice, there will always be additive noise as separate speech signals in the presence of Gaussian well as reverberations in an acoustic environment that noise, even occasionally resulting in a reduced signal- make obtaining an ideal mixed signal difficult if not to-noise ratio (SNR) relative to the original mized sig- impossible. The second issue regarding ICA is its ap n&. The technique proposed here utilizes a kurtosis- plication to situations where a large number of micro- based genetic microphone selection mechanism in order phones are present, but only a few independent speech to select an optimal set of microphones from a large sources are present. Take, for example, a small room number of microphones that are agsumed to be wad- with 20 microphones and three simultaneous speakers. able. ICA is performed on the sign& of the selected Does one perform ICA on all 20 microphones, which is microphones and the avemge output kurtosis is used to computationally expensive, or should a certain set of update the microphone selections. This technique re- microphones, say 3, be selected and used? If a small sults in an average improvement of approzimately 11 set of microphones is selected, then how is this selec- dB in the output SNR relative to the result obtained by tion accomplished? What metric can be used to aid a random selection of microphones. The dmwback of the selection process? this technique is its failure in the presence of leptokur- T&~ paper attempts to ansmr these questions by tic noise. proposing a separation-quality metric based on the kurtosis of each of the separated channels. This metric