Article
Investigation of VoicePrivacy Challenge Baseline B1 performance under different noise conditions (en)
* Presenting author
Abstract:
Speaker anonymization is a relatively new field, where the aim is to minimize the identity information leakage in voice recordings while maintaining the usefulness for other downstream tasks such as automatic speech recognition. The VoicePrivacy Challenge (VPC) framework, which provides training and evaluation datasets, evaluation metrics and baselines, is commonly utilized in the publications. However, the adopted evaluation framework contains only clean speech recordings. In this work, we investigate the robustness of the VPC Baseline B1 under different noise conditions including reverberation, babble noise, and microphone noise. We measure and report the deviations from the VoicePrivacy evaluation metrics.