Vocal Tract Length Perturbation-based Pseudo-Speaker Augmentation Considering Speaker Variability for Speaker Verification

In this paper, we propose an advanced method for pseudo-speaker augmentation using vocal tract length perturbation (VTLP) for automatic speaker verification (ASV) systems. The state-of-the-art ASV systems based on speaker embeddings require a substantial amount of training data to construct a reliable speaker embedding extractor. Traditional data augmentation methods for ASV typically focus on increasing the corpus size, while ensuring sufficient diversity of distinct speakers is also crucial for improving accuracy. A previous study has reported that VTLP is used as an effective pseudo-speaker generation method, and increasing the number of speakers through VTLP can enhance ASV performance. However, the previous method has demonstrated limitations in the number of pseudo-speakers that can be effectively used, indicating that these methods may not be sufficiently effective. Therefore, this paper proposes increasing the number of pseudo-speakers available for data augmentation by setting the VTLP parameters to ensure diversity for each speaker. The experimental results show that the proposed pseudo-speaker augmentation method can significantly improve the performance of ASV system based on ECAPA-TDNN.