Aiming at the problems of high computing resource consumption caused by wav2vec2.0 model in existing detection sys-tems and the limited generalization performance of traditional scoring methods,a spoofing detection algorithm based on shallow feature fusion of clustering center was proposed to solve the above problems.The deep layer of wav2vec2.0 model was trimmed,and the shallow features were pooled by attention mechanism to shorten the time series length.The linear layer was used to determine the fusion weight.The clustering centers were obtained by K-means++,and the representation cosine similarity between the current sample and the corresponding class center was adapted for training and scoring to distinguish bona-fide and spoofing speech.Results of experiment on the datasets of ASVspoof2019 and ASVspoof2021 challenges of the logical track show that the scale of wav2vec2.0 model parameter is reduced by 60％,and the equal error rate reaches 0.34％and 3.67％.It is sig-nificantly better than the similar wav2vec2.0 frond-end model and the traditional scoring method of classifiers in terms of model simplification and generalization performance.
维普
万方数据