Input 1
Original
spk1
spk2
spk3
spk4
Interspeech 2026 · Audio samples
Acoustic token admixture for joint speaker and content anonymization
Our system hides two things in a recording: who is speaking (biometric identity) and what identifies them (named entities). It works on a single stream of RVQ-Whisper acoustic tokens. Encoder-derived tokens and phoneme-conditioned tokens are mixed frame by frame, with mixing ratio β and a cosine-similarity gate τ = 0.6. When NER flags a sensitive span, the system regenerates only those frames and leaves the surrounding prosody untouched. On VoicePrivacy 2024 it reaches EER 42.54% with WER 3.73% at β = 0.7.
All anonymized outputs below are 16 kHz BigVGAN vocoder outputs. Use headphones for the fairest comparison.
Section 1
Anonymization and speaker consistency
Each utterance is anonymized with four pseudospeakers, each fixed by a random seed (spk1 to spk4 are seeds 3358, 4930, 5192 and 5978). Listen down a column to hear one pseudospeaker stay consistent across different content. Listen along a row to hear the same content rendered by different pseudospeakers.
β = 0.7
Columns: Original · spk1 · spk2 · spk3 · spk4
Input 2
Original
spk1
spk2
spk3
spk4
Input 3
Original
spk1
spk2
spk3
spk4
Input 4
Original
spk1
spk2
spk3
spk4
Section 2
Speech editing
NER-triggered content replacement (paper, Sec. 2.3). orig is the source recording and anon is the anonymized version with unchanged content. v1 and v2 replace the detected entity with a different name or place. Only the frames of the entity span are regenerated. They are spliced back with an 8-frame (about 160 ms) crossfade, so the pseudospeaker and the surrounding prosody stay intact.
β = 0.7 · NER span replacement
Sample 1
“The American planes maneuvered in such a way as to encircle the Taube.”
- v1: American → Australian
- v2: American → English
orig
anon
v1
v2
Sample 2
“As they sat down, a burst of applause drew Alexander's attention to the stage.”
- v1: Alexander → William
- v2: Alexander → Jonathan
orig
anon
v1
v2
Sample 3
“The power and enterprise of Spain were directed to wider and richer fields.”
- v1: Spain → France
- v2: Spain → Germany
orig
anon
v1
v2
Sample 4
“However, Jaime Alguersuari failed to score any points in the season.”
- v1: Jaime Alguersuari → David Jones
- v2: Jaime Alguersuari → Victor Smith
orig
anon
v1
v2
Section 3
Admixture effect
The same utterance anonymized at β = 0.0, 0.7 and 1.0. β is the probability that a frame uses a phoneme-derived token instead of the encoder-derived token. At β = 0.0 the only change is the pseudospeaker substitution, so more of the source speaking style survives. Higher β removes more residual speaker cues from the token stream. On VoicePrivacy 2024, EER rises from 23.0% at β = 0.0 to 42.54% at β = 0.7. β = 1.0 is included to hear the fully phoneme-derived end of the range; the paper does not report it.
Sample 1
orig
β = 0.0
β = 0.7
β = 1.0
Sample 2
orig
β = 0.0
β = 0.7
β = 1.0
Sample 3
orig
β = 0.0
β = 0.7
β = 1.0
Sample 4
orig
β = 0.0
β = 0.7
β = 1.0