Research Of Speaker Recognition Based On Ant Colony Optimization Algorithm

Speaker recognition deals with recognizing the identity of the person speaking utterance; it is the process of automatically recognizing who is speaking based on the information obtained from the speech. Speaker recognition has a wide range of applications which include banking or credit card transitions by telephone, information and reservation services, access control in high security areas and forensic investigations.Though speaker recognition system perform well when clean speech is used for training and testing, the performance degrades rapidly when speech used in real-world conditions. The focus of this research effort is to develop techniques for training GMM parameters.Ant Colony Optimization (ACO) is a random search algorithm. Similar to the other evolutionary algorithms, it finds the optimal solution based on the evolutionary process of candidate group of searching paths. ACO speeds up the evolutionary process by positive feedback of pheromone. It is also a parallel algorithm. Through communication and transferring among the ants, they could collaborate with each other for finding the best path. ACO is widely applied on Traveling Salesman Problem (TSP) and other optimization problems.After analyzing the current researches on speaker recognition and Ant Colony Optimization (ACO) algorithm, in this thesis, introducing the ACO to the parameter training process of Gaussian mixture model for building speaker recognition system was proposed. Also the TopN method for recognition comparing based on eigenvectors was proposed. A speaker recognition system was realized based on our approaches.The main contributions in the thesis are as follows:1. The current research works on speaker recognition system and its basic structure and the assessment criteria of recognition correctness was discussed.2. The ACO algorithm was analyzed and the theories and characters of ACO and Gaussian mixture model were studied. Then introducing ACO to the parameter training process of Gaussian mixture model was proposed. Given the speech signal of speakers, the model parameters were assessed by clustering the signals using ACO. Finally, the models were built for each speaker. 3. Considering the limited number of training data and variation of speech signals, the TopN method which only uses the most N% similar eigenvectors for recognition comparing was proposed. By using this method, the effect of useless speech signals to the whole signals was significantly reduced.4. Based on the above proposals, a speaker recognition system was realized with Visual C++. The system has the ability of recording the speeches, training, recognition, etc. The experimental results show that, the ACO based speaker recognition system has high recognition rate. It is a new and effective method for speaker recognition.