Universal Semantic Disentangled Privacy-preserving Speech Representation Learning¶
🔴 Supplementary material [Anonymous submission ICLR 2025]¶
A more elaborate non-anonymous sample page will be provided to the audience upon ICLR de-anonymization, showcasing the examples presented on this sample page.
We present samples for two voices from our internal voice repository: one female voice and one male voice.
Abstract: The use of audio recordings of human speech to train LLMs poses privacy concerns due to these models' potential to generate outputs that closely resemble artifacts in the training data. In this study, we propose a speaker privacy-preserving representation learning method through the Universal Speech Codec (USC), a computationally efficient encoder-decoder model that disentangles speech into: (i) privacy-preserving semantically rich representations, capturing content and speech paralinguistics, and (ii) residual acoustic and speaker representations that enables high-fidelity reconstruction. Extensive evaluations presented show that USC's semantic representation preserves content, prosody, and sentiment, while removing potentially identifiable speaker attributes. Combining both representations, USC achieves state-of-the-art speech reconstruction. Additionally, we introduce an evaluation methodology for measuring privacy-preserving properties, aligning with perceptual tests. We compare USC against other codecs in the literature and demonstrate its effectiveness on privacy-preserving representation learning, illustrating the trade-offs of speaker anonymization, paralinguistics retention and content preservation in the learned semantic representations.
Universal Speech Codec¶
High-fidelity Reconstruction¶
Utterances | Recordings (24 kHz) |
EnCodec C0:7 6.0 kbps (24 kHz) |
DAC C0:8 7.75 kbps (44.1 kHz) |
SpeechTokenizer C0:7 4.00 kbps (16 kHz) |
FaCodec C0:5 4.80 kbps (16 kHz) |
USC C0:5 1.6 kbps (24 kHz) |
---|---|---|---|---|---|---|
Sample 1 | ||||||
Sample 2 | ||||||
Sample 3 | ||||||
Sample 4 | ||||||
Sample 5 | ||||||
Sample 6 |
Semantic Reconstruction¶
Utterances | Recordings (24 kHz) |
EnCodec C0 0.75 kbps (24 kHz) |
DAC C0 0.86 kbps (44.1 kHz) |
SpeechTokenizer C0 0.50 kbps (16 kHz) |
FaCodec C0:2 2.40 kbps (16 kHz) |
USC C0 0.35 kbps (24 kHz) |
---|---|---|---|---|---|---|
Sample 1 | ||||||
Sample 2 | ||||||
Sample 3 | ||||||
Sample 4 | ||||||
Sample 5 | ||||||
Sample 6 |
Voice Conversion through semantic Partial-Teacher-Forcing (PTF)¶
Female to Male speaker conversion¶
Utterances | Source speaker (16 kHz) |
Target speaker (16 kHz) |
PTF + USC (24 kHz) |
---|---|---|---|
Sample 1 | |||
Sample 2 | |||
Sample 3 | |||
Sample 4 | |||
Sample 5 |
Male to Female speaker conversion¶
Utterances | Source speaker (16 kHz) |
Target speaker (16 kHz) |
PTF + USC (24 kHz) |
---|---|---|---|
Sample 1 | |||
Sample 2 | |||
Sample 3 | |||
Sample 4 | |||
Sample 5 |