Speaker identification using hybrid of subtractive clustering and radial basis function

摘要

Speaker identification is the computing task of identifying unknown identities based on voice. A good speaker identification system must have a high accuracy rate to prevent incorrect detection of the user's identity. This research proposed a hybrid of Subtractive Clustering and Radial Basis Function (Sub-RBF) which is the combination of supervised and unsupervised learning. Unsupervised learning is more suitable for learning large and complex models than supervised learning. This is because supervised learning increasing the number of connections between sets in the network. If the model contains a large and complex dataset, supervised learning is difficult. In addition, K-means is faced with improper initial guessing of first cluster centre and difficulty in determining the number of cluster centres. The proposed technique is introduced because subtractive clustering is able to solve these problems. RBF is a simple network structures with fast learning algorithm. RBF neural network model with subtractive clustering proposed to select hidden node centers, can achieve faster training speed. In the meantime, the RBF network was trained with a regularization parameter so as to minimize the variances of the nodes in the hidden layer and perform more accurate prediction. The accuracy rate for subtractive clustering is 8.125% and 11.25% for training dataset 1 and training dataset 2 respectively. However, Sub-RBF provides 76.875% and 71.25% accuracy rate for training dataset 1 and training dataset 2 respectively. In conclusion, Sub-RBF has improved the speaker identification system accuracy rate.