Time-Frequency Masking: Linking Blind Source Separation and Robust Speech Recognition
Marco Khne, Roberto B. Togneri, Sven Erik Nordholm · InTech eBooks · 2008
The experimental results reported here suggest that DUET might be used as an effective front-end for missing data speech recognition. Its simplicity, robustness and easy integration into existing ASR architecture are the main compelling arguments for the proposed model. It also fundamentally differs from other multi-channel approaches in the way it makes use of spatial information. Instead of filtering the corrupted signal to retrieve the sources (McCowan et al., 2000; Low et al., 2004, Seltzer et al., 2004a) the time-frequency plane is partitioned into disjoint regions each assigned to a particular source. A key aspect of the model is the histogram peak detection. Here, we assumed prior knowledge about the number of speakers which should equal the number of peaks in the histogram. However, for a high number of simultaneous speakers the sparseness assumption becomes increasingly unrealistic and as a consequence sometimes histogram peaks are not pronounced enough in the data set. Forcing the peak detection algorithm to find an inadequate number of peaks will produce false localization results. Ultimately, the algorithm should be able to automatically detect the number of sources visible in the data which is usually denoted as unsupervised clustering. This would indeed make the source separation more autonomous and truly blind. However, unsupervised clustering is a considerably more difficult problem and is still an active field of research (Grira et al., 2004). Other attempts to directly cluster the attenuation and delay distributions using a statistical framework have been reported elsewhere (Araki et al., 2007; Mandel et al., 2006) and would lead to probabilistic mask interpretations. that was kept very small to avoid phase A point of concern is the microphone distance ambiguities (Yilmaz & Rickard, 2004). Clearly, this limits the influence of the attenuation parameter (see Fig. 2a). Rickard (2007) has offered two extensions to overcome the small sensor spacing by using phase differentials or tiled histograms. Another option to consider