PDF-NET: Pitch-adaptive Dynamic Filter Network for Intra-gender Speaker Verification

In this paper, we propose a method to significantly improve the performance of the intra-gender comparison task in a neural speaker verification (SV) system. Recent state-of-the-art neural SV models have a substantial performance decline in the same gender comparison trials, but few studies have comprehensively investigated this problem. To this end, we develop a light-weight pitch-adaptive dynamic filter network (PDF-NET) that generates the convolutional filters conditioned on the input pitch information, from the observation that intra-gender speakers share a similar pitch range. By doing so, speaker embeddings from the same speakers are well clustered with each other even within the same gender class. Experimental results show that our proposed method outperforms the baseline for various SV datasets. In particular, female SV trials on LibriSpeech performed herein achieve a relative improvement of 11.4% in terms of equal error rate with only 1.6k additional parameters.