ASF-Conformer: Audio Scoring Conformer with FFC for Speaker Verification in Noisy Environments

Background noise significantly impacts speech intelligibility, reducing the accuracy and reliability of the speaker verification system. Most existing noise reduction algorithms are specific to certain types of noise and have limitations, making them ineffective on eliminating background noise. Therefore, the extraction of robust features and the development of noise-resistant models that adapt to various noisy environments remain crucial challenges in the field of speaker verification. In this paper, we propose a Conformer-based Audio Scoring Conformer with Fast Fourier Convolution (ASF-Conformer), which is a speaker verification model. Firstly, the audio scoring module is introduced to evaluate and weight the audio features, aiming to select more robust features in noisy environments. Secondly, we introduce Fast Fourier Convolution as a replacement for the Conformer’s convolution module, improving the model’s ability to capture global features while reducing the model parameters. Finally, this paper conducts comparative tests with the current mainstream models on public dataset VoxCeleb1, and synthesized noisy dataset Mu-VoxCeleb1. The experimental results demonstrate that the proposed ASF-Conformer model, compared to the ECAPA-TDNN model with essentially the same parameters, outperforms ECAPA-TDNN by 2% and 18% respectively when evaluated using the EER metrics on the VoxCeleb1 and Mu-VoxCeleb1 datasets. These results highlight the effectiveness of the proposed model in enhancing the accuracy of speaker verification tasks, especially in noisy environments.