Robust speaker verification method for VoIP-transmitted utterances using HiFi-GAN

Speaker verification systems are increasingly handle utterances transmitted via internet applications. The performance of speaker verification systems is known to degrade under mismatched conditions, where enrollment is microphone-recorded utterance and verification is utterance transmitted over the internet, which are distorted by codec compression and packet loss. To address this issue, this paper proposes a speaker verification method that utilizes a HiFi-GAN to suppress performance degradation. The proposed method converts the internet-transmitted utterance into the original microphone recording. This is achieved by employing a HiFi-GAN model fine-tunedon microphone-recorded utterances. For our experiments, we re-recorded the utterances in Japanese Newspaper Article Sentences (JNAS) corpus through three VoIP applications: Microsoft Teams, Skype, and Discord. In our experiment, we used the ResNet based speaker verification system. The results first confirmed a significant deterioration in speaker verification accuracy for internet-transmitted utterances, with the Equal Error Rate (EER) increasing from 2.97% for the microphone-recording utterances to 23.81% (Skype), 28.78% (Discord), and 22.81% (Teams). Furthermore, the experiments demonstrated that the proposed method effectively suppressed these degradations, reducing the EERs for the three applications to 21.51% (Skype), 17.46% (Discord), and 20.75% (Teams), respectively. [Work supported by JSPS KAKENHI Grant Numbers JP23K11165, JP25K15125, JP24K14988.]