Optimal fusion of diverse feature sets for speaker identification: an alternative method

For speaker identification, a robust and effective feature extraction method is necessary. But in the current circumstance, there exists no perfect feature that could optimally characterize physiological difference among speakers regardless of personal variation. A soft competition scheme for optimal fusion of diverse feature sets is applied to speaker identification in order to achieve the improved performance. Based on a linear combination scheme, diverse feature vectors are used together while the winning feature vector through soft competition play s more important role in the representation. The simulations on KING corpus show that this alternative method could yield good performance for speaker identification.