Research

Paper

AI LLM March 09, 2026

Privacy-Preserving End-to-End Full-Duplex Speech Dialogue Models

Authors

Nikita Kuzmin, Tao Zhong, Jiajun Deng, Yingke Zhu, Tristan Tsoi, Tianxiang Cao, Simon Lui, Kong Aik Lee, Eng Siong Chng

Abstract

End-to-end full-duplex speech models feed user audio through an always-on LLM backbone, yet the speaker privacy implications of their hidden representations remain unexamined. Following the VoicePrivacy 2024 protocol with a lazy-informed attacker, we show that the hidden states of SALM-Duplex and Moshi leak substantial speaker identity across all transformer layers. Layer-wise and turn-wise analyses reveal that leakage persists across all layers, with SALM-Duplex showing stronger leakage in early layers while Moshi leaks uniformly, and that Linkability rises sharply within the first few turns. We propose two streaming anonymization setups using Stream-Voice-Anon: a waveform-level front-end (Anon-W2W) and a feature-domain replacement (Anon-W2F). Anon-W2F raises EER by over 3.5x relative to the discrete encoder baseline (11.2% to 41.0%), approaching the 50% random-chance ceiling, while Anon-W2W retains 78-93% of baseline sBERT across setups with sub-second response latency (FRL under 0.8 s).

Metadata

arXiv ID: 2603.08179
Provider: ARXIV
Primary Category: eess.AS
Published: 2026-03-09
Fetched: 2026-03-10 05:43

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.08179v1</id>\n    <title>Privacy-Preserving End-to-End Full-Duplex Speech Dialogue Models</title>\n    <updated>2026-03-09T10:01:24Z</updated>\n    <link href='https://arxiv.org/abs/2603.08179v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.08179v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>End-to-end full-duplex speech models feed user audio through an always-on LLM backbone, yet the speaker privacy implications of their hidden representations remain unexamined. Following the VoicePrivacy 2024 protocol with a lazy-informed attacker, we show that the hidden states of SALM-Duplex and Moshi leak substantial speaker identity across all transformer layers. Layer-wise and turn-wise analyses reveal that leakage persists across all layers, with SALM-Duplex showing stronger leakage in early layers while Moshi leaks uniformly, and that Linkability rises sharply within the first few turns. We propose two streaming anonymization setups using Stream-Voice-Anon: a waveform-level front-end (Anon-W2W) and a feature-domain replacement (Anon-W2F). Anon-W2F raises EER by over 3.5x relative to the discrete encoder baseline (11.2% to 41.0%), approaching the 50% random-chance ceiling, while Anon-W2W retains 78-93% of baseline sBERT across setups with sub-second response latency (FRL under 0.8 s).</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='eess.AS'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='eess.SP'/>\n    <published>2026-03-09T10:01:24Z</published>\n    <arxiv:primary_category term='eess.AS'/>\n    <author>\n      <name>Nikita Kuzmin</name>\n    </author>\n    <author>\n      <name>Tao Zhong</name>\n    </author>\n    <author>\n      <name>Jiajun Deng</name>\n    </author>\n    <author>\n      <name>Yingke Zhu</name>\n    </author>\n    <author>\n      <name>Tristan Tsoi</name>\n    </author>\n    <author>\n      <name>Tianxiang Cao</name>\n    </author>\n    <author>\n      <name>Simon Lui</name>\n    </author>\n    <author>\n      <name>Kong Aik Lee</name>\n    </author>\n    <author>\n      <name>Eng Siong Chng</name>\n    </author>\n  </entry>"
}