研究目的
To present an approach to calibrating the output scores of a speaker verification system using the time interval between comparison samples as additional information to compensate for the detrimental impact of aging on speaker verification performance.
研究成果
The proposed score-aging calibration was shown to provide better performance than by simply pooling all of the data to train conventional calibration parameters. The approach can extrapolate well to unseen time differences, representing a straightforward means of compensating for the detrimental impact of aging on speaker verification performance.
研究不足
The TCDSA corpus involves a very small number of speakers (17), a sparsity of trials, and inherent recording variability due to the large time intervals present in the data, which limits the conclusions that can be drawn from the TCDSA experiments.
1：Experimental Design and Method Selection
The study extends conventional score calibration by incorporating aging information as the time difference between the corresponding trials. Several functions are proposed for the incorporation of this time information in a conventional linear score calibration transformation.
2：Sample Selection and Data Sources
Two corpora are used for the experiments: the Multi-session Audio Research Project (MARP) corpus for short-term aging data and the Trinity College Dublin Speaker Aging (TCDSA) corpus for long-term aging data.
3：List of Experimental Equipment and Materials
An i-vector framework with probabilistic linear discriminant analysis (PLDA) modeling is used for the speaker verification experiments. Mel Frequency Cepstral Coefficients (MFCCs) of 13 dimensions were extracted over 20 ms windows at 10 ms intervals.
4：Experimental Procedures and Operational Workflow
A cross-validation experiment was applied, whereby the average performance of multiple development and test set divisions is used to compare conventional and score-aging calibration methods.
5：Data Analysis Methods
Several metrics are used to compare the discrimination and calibration performance of conventional and aging calibration approaches, including Equal Error Rate (EER), Cost of Log-Likelihood Ratio (C(cid:2)(cid:2)r), and Minimum Cost of Log-Likelihood Ratio (C m in (cid:2)(cid:2)r).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
