Disentanglement-based speaker anonymization involves decomposing speech into a semantically meaningful representation, altering the speaker embedding, and resynthesizing a waveform using a neural vocoder. State-of-the-art systems of this kind are known to remove emotion information. Possible reasons include mode collapse in GAN-based vocoders, unintended modeling and modification of emotions through speaker embeddings, or excessive sanitization of the intermediate representation. In this paper, we conduct a comprehensive evaluation of a state-of-the-art speaker anonymization system to understand the underlying causes. We conclude that the main reason is the lack of emotion-related information in the intermediate representation. The speaker embeddings also have a high impact, if they are learned in a generative context. The vocoder's out-of-distribution performance has a smaller impact. Additionally, we discovered that synthesis artifacts increase spectral kurtosis, biasing emotion recognition evaluation towards classifying utterances as angry. Therefore, we conclude that reporting unweighted average recall alone for emotion recognition performance is suboptimal.
View on arXiv