Research on Front-end of ASV System Based on Mel Spectrum in Noise Scenario

Mel Spectrum is an engineered audio feature that simulates human perception and has been successfully used in various speech and audio applications, making it a versatile frame-level feature. Nevertheless, the Mel Spectrum still needs further improvement. This paper proposes a frontend architecture for ASV systems based on deep neural networks (DNNs) and an improved Mel spectrum. Our architecture retains the simplicity and interpretability based on the characteristics of the Mel spectrum while providing noise immunity and cross-domain robustness. In this paper, we use the output of the studied acoustic feature extractor as the input to the ECAP A- TDNN and complete the training of an ECAPA-TDNN model for 1211 speaker identifications based on the VoxCeleb1-dev dataset. An automatic speaker verification (ASV) system was constructed using a model as a speaker embedding extractor. The experimental results show that the two ASV systems containing the improved Mel spectrum proposed in this paper show significant improvements compared to the ASV systems containing MFCCs in both intra-domain and cross-domain scenarios. On VoxCeleb1-dev and VoxMovies, the EER of the ASV system containing the improved Mel spectrum features––Mel-MACs achieves an improvement of 4.34% and 19.81 % compared to the ASV system containing FBanks and MFCCs, respectively.