Research On Speaker Recognition Based On Integrated Learning

With the rapid development of various bionic intelligence technologies,speaker recognition systems applicable to specific scenarios have emerged.As an important research branch of biomimetic feature recognition,speaker recognition technology has been widely used in various fields due to its wide versatility,long-lasting stability and low cost.In view of its powerful nonlinear classification capability,the ensemble learning algorithms have been applied to speaker recognition systems and have achieved some successes,however the traditional ensemble learning algorithm still suffers from insufficient recognition accuracy,low efficiency,poor robustness and weak classification capability for unbalanced data.This paper addressed the shortcomings of traditional ensemble learning algorithms,conducted in-depth research around the needs of multi-scene speech recognition,and proposed ensemble learning classification algorithms based on bioimmunology theory for a variety of practical situations,which promoted the application of speaker recognition technology in criminal investigation,identity confirmation,security and other fields.The main research work and innovation points of this paper are as follows.(1)The Bal-Adaboost(Balance-Adaboost)algorithm is proposed for the problem of Costly misclassification and low accuracy when traditional ensemble learning algorithms deal with unbalanced data and applied to speaker gender classification.Inspired by the classifier evaluation criteria,the algorithm adds an adaptively adjustable penalty term to the traditional Adaboost algorithm,and effectively solves the imbalanced data classification problem by achieving optimal data division.The algorithm is validated based on the open unbalanced dataset and the speaker gender dataset,and the results show that the Bal-Adaboost algorithm exhibits higher accuracy on both types of datasets and can achieve the optimal classification of the data to be classified.(2)The Immune Boost(Immune-Adaboost)algorithm based on bioimmune characteristics is proposed and applied to single-class speaker recognition to address the problems of slow classification speed and poor robustness of traditional ensemble learning algorithms.The algorithm draws on the bioimmune property and utilizes the synergy of innate immune classifier and adaptive immune classifier to solve the problems of high algorithm complexity,long time consuming and high training error of traditional Adaboost algorithm.In this paper,we compare the performance of the algorithm in different data sets through numerous experiments.The results show that the Immune Boost algorithm exhibits high classification accuracy,robustness and high efficiency on single-class speaker voiceprint data sets,Gaussian data sets,and open datasets.(3)The MC-Immune Boost(Multi Class-Immune Boost)algorithm is proposed and applied to multi-class speaker recognition to address the problems of low accuracy and slow operation of traditional ensemble learning algorithms for multi-class classification tasks.The algorithm takes the decision strategy of the innate immune classifier as the starting point and proposes an MC-Immune classifier based on an improved decision criterion,which further improves the effectiveness of the MC-Adaptive Immune classifier.The MC-Immune Boost algorithm proposed in this paper achieves good recognition performance on both multi-class speaker voiceprint datasets and open datasets.