Abstract: In this paper, a novel approach for speaker identification called Fishervoice is proposed. It was inspired by the two-dimensional (2D) Fisherface technique, which is a method that combines a two-stage ``PCA+LDA'' strategy and two-dimensional discrimination techniques. Experimental results on the BANCA database demonstrate that the Fishervoice approach is effective for speaker identification tasks, in particular when there are mismatched conditions. It is remarkable the reduction on the number of parameters needed to describe each speaker model achieved with the Fishervoice technique, thereby causing a reduction of computational and memory costs.
Index Terms: fishervoices, speaker identification, dimensionality reduction.