Ensemble learning (EL)-based speaker verification method

An ensemble learning (EL)-based speaker verification method, comprising the following steps:

• step 1:
• performing data acquisition and preprocessing;
• step 2:
• selecting and training a group of basic models, and optimizing model parameters;
• and
• step 3:
• constructing an ensemble model, comprising the following steps:
• step 3.1:
• converting acoustic feature vectors of an input pair of voice samples into a pair of speaker feature embedding via the group of basic models, and scoring the voice samples based on a similarity to separately obtain similarity scores of the group of basic models;
• step 3.2:
• constructing a detection cost function (DCF);
• and
• step 3.3:
• generating a weight;
• performing weighted fusion on scoring results of the group of basic models based on the weight to obtain a total similarity score list;
• traversing score values in the total score list one by one as a current temporary threshold, wherein if a score value is greater than the current temporary threshold, it is determined that there is a same speaker, if the score value is not greater than the temporary thresholds, it is determined that there are different speakers, calculating correctness rates of training set data under values of all temporary thresholds, and taking a temporary threshold with a highest correctness rate as a threshold;
• based on the threshold, determining, based on the total similarity score list, a positive number of scores greater than the threshold and a negative number of scores less than the threshold;
• calculating a target detection result by using the DCF;
• generating a new weight, and repeating the foregoing operation to obtain a new target detection result;
• and based on a difference of the target detection result, iterating a weight until an iteration ends, to obtain a final ensemble model for speaker verification.