Evaluation Metrics

1. Speaker Recognition Accuracy

In speaker recognition, there are two tasks: automatic speaker identification (ASI) and automatic speaker verification (ASV). The ASI task can be directly evaluated using the identification accuracy (Accuracy, ACC) on the test set, while the ASV task requires calculating the Equal Error Rate (EER) using the trials list.

1.1 Accuracy

In the speaker identification task, ACC is typically calculated as the ratio of correctly identified speakers to the total number of test samples. This metric intuitively reflects the system's ability to correctly recognize different speakers. The higher the value, the better the identification accuracy. ACC can be used to compare the performance of different speaker identification systems, providing a basis for selecting a more efficient system.

1.2 Equal Error Rate

In the speaker verification task, the EER is calculated as follows: as the system's decision threshold changes, different False Accept Rates (FAR) and False Reject Rates (FRR) are obtained. The EER is the error rate when FAR and FRR are equal. EER is an important metric for evaluating the performance of speaker verification systems, with lower values indicating better system performance.