Learning De-identified Representations of Prosody from Raw Audio
This work addresses the need for privacy-preserving speech processing by developing a method to extract prosody without revealing speaker identity, which is incremental as it builds on prior contrastive self-supervised approaches but introduces novel biases for de-identification.
The paper tackles the problem of learning de-identified prosody representations from raw audio by introducing inductive biases to minimize timbral information and decouple prosody from speaker representations, achieving performance comparable to state-of-the-art speech representations on the DAMMP benchmark and demonstrating reduced identifiability.
We propose a method for learning de-identified prosody representations from raw audio using a contrastive self-supervised signal. Whereas prior work has relied on conditioning models on bottlenecks, we introduce a set of inductive biases that exploit the natural structure of prosody to minimize timbral information and decouple prosody from speaker representations. Despite aggressive downsampling of the input and having no access to linguistic information, our model performs comparably to state-of-the-art speech representations on DAMMP, a new benchmark we introduce for spoken language understanding. We use minimum description length probing to show that our representations have selectively learned the subcomponents of non-timbral prosody, and that the product quantizer naturally disentangles them without using bottlenecks. We derive an information-theoretic definition of speech de-identifiability and use it to demonstrate that our prosody representations are less identifiable than other speech representations.