Robust Speaker Recognition for Whispered Speech
Nikita Khmelev, Anastasia Avdeeva, Sergey Arkad'yevich Novoselov, Artem Chirkovskiy, Marina Volkova · 2025
Although significant progress has been made in developing stable and robust speaker verification systems, they still have some limitations. When a speaker's style changes, it can affect the performance of the verification system. In this paper, we consider the whispered speech domain. In particular, the adaptation of the verification system based on wav2vec 2.0 is proposed in order to achieve high quality in both the whispered and normal speech domains. Augmentation techniques based on whisper synthesis using Praat software, along with a domain score normalization approach, are utilized to improve performance in the whispered domain. Additionally, a method for whisper detection is employed. The proposed approaches significantly enhance system performance on whispered data.