Speaker information modification in the VoicePrivacy 2020 toolchain

This paper presents a study of the baseline system of the VoicePrivacy 2020 challenge. This baseline relies on a voice conversion system that aims at separating speaker identity and linguistic contents for a given speech utterance. To generate an anonymized speech waveform, the neural acoustic model and neural waveform model use the related linguistic content together with a selected pseudo-speaker identity. The linguistic content is estimated using bottleneck features extracted from a triphone classifier while the speaker information is extracted then modified to target a pseudo-speaker identity in the x-vector's space. In this work, we first proposed to replace the triphone-based bottleneck features extractor that requires supervised training by an end-to-end Automatic Speech Recognition (ASR) system. In this framework, we explored the use of adver-sarial and semi-adversarial training to learn linguistic features while masking speaker information. Last, we explored several anonymization schemes to introspect which module benefits the most from the generated pseudo-speaker identities.