Investigation of the influence of blocks on the linear spectrum for synthetic speech detection

Abstract

AbstractIn recent years, constant‐Q cepstral coefficient (CQCC) has successfully used in synthetic speech detection (SSD), however, linear‐domain logarithm power spectrum (LLPS) information is not fully captured by discrete cosine transform (DCT) in CQCC extraction and how the block affects the LLPS information extraction for SSD has not been investigated in the previous studies. In order to investigate how the block affects the LLPS information extraction and extract more detailed information from LLPS for SSD, a new feature, constant‐Q block coefficient (CQBC), is proposed by modifying CQCC using block transform plus DCT on LLPS in this letter. Furthermore, how the length of block affects the performance is deeply investigated in this work. The experimental result on ASVspoof 2015 evaluation set indicates that: (1) there is much difference in the performance for different length block (2) CQBC‐DA can improve the performance of SSD and its average equal error rate can reach 0.0724, which decline 36% compared with CQCC‐DA (3) CQBC‐DA is superior to many front‐ends that have been benchmarked in terms of average equal error rate.