Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Content Anonymization for Privacy in Long-form Audio

Cristina Aggazzotti, Ashi Garg, Zexin Cai, Nicholas Andrews

TL;DR

The paper addresses privacy risks in long-form audio where speaker identity can be inferred from linguistic content despite voice anonymization. It introduces a content anonymization method that rewrites transcripts within an ASR-TTS pipeline using contextual paraphrasing to erase speaker-style cues while preserving meaning. Through experiments on the Fisher corpus, the authors show that content-based attacks become more powerful with more utterances, but paraphrasing—especially segment-based—substantially mitigates this risk with acceptable utility. They demonstrate that full voice+content anonymization yields privacy gains with natural-sounding output and preserved semantic content, suggesting practical adoption for long-form audio.

Abstract

Voice anonymization techniques have been found to successfully obscure a speaker's acoustic identity in short, isolated utterances in benchmarks such as the VoicePrivacy Challenge. In practice, however, utterances seldom occur in isolation: long-form audio is commonplace in domains such as interviews, phone calls, and meetings. In these cases, many utterances from the same speaker are available, which pose a significantly greater privacy risk: given multiple utterances from the same speaker, an attacker could exploit an individual's vocabulary, syntax, and turns of phrase to re-identify them, even when their voice is completely disguised. To address this risk, we propose new content anonymization approaches. Our approach performs a contextual rewriting of the transcripts in an ASR-TTS pipeline to eliminate speaker-specific style while preserving meaning. We present results in a long-form telephone conversation setting demonstrating the effectiveness of a content-based attack on voice-anonymized speech. Then we show how the proposed content-based anonymization methods can mitigate this risk while preserving speech utility. Overall, we find that paraphrasing is an effective defense against content-based attacks and recommend that stakeholders adopt this step to ensure anonymity in long-form audio.

Content Anonymization for Privacy in Long-form Audio

TL;DR

The paper addresses privacy risks in long-form audio where speaker identity can be inferred from linguistic content despite voice anonymization. It introduces a content anonymization method that rewrites transcripts within an ASR-TTS pipeline using contextual paraphrasing to erase speaker-style cues while preserving meaning. Through experiments on the Fisher corpus, the authors show that content-based attacks become more powerful with more utterances, but paraphrasing—especially segment-based—substantially mitigates this risk with acceptable utility. They demonstrate that full voice+content anonymization yields privacy gains with natural-sounding output and preserved semantic content, suggesting practical adoption for long-form audio.

Abstract

Voice anonymization techniques have been found to successfully obscure a speaker's acoustic identity in short, isolated utterances in benchmarks such as the VoicePrivacy Challenge. In practice, however, utterances seldom occur in isolation: long-form audio is commonplace in domains such as interviews, phone calls, and meetings. In these cases, many utterances from the same speaker are available, which pose a significantly greater privacy risk: given multiple utterances from the same speaker, an attacker could exploit an individual's vocabulary, syntax, and turns of phrase to re-identify them, even when their voice is completely disguised. To address this risk, we propose new content anonymization approaches. Our approach performs a contextual rewriting of the transcripts in an ASR-TTS pipeline to eliminate speaker-specific style while preserving meaning. We present results in a long-form telephone conversation setting demonstrating the effectiveness of a content-based attack on voice-anonymized speech. Then we show how the proposed content-based anonymization methods can mitigate this risk while preserving speech utility. Overall, we find that paraphrasing is an effective defense against content-based attacks and recommend that stakeholders adopt this step to ensure anonymity in long-form audio.
Paper Structure (13 sections, 3 figures, 1 table)

This paper contains 13 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Methodology
  5. Experiments
  6. Experimental Details
  7. Attack models
  8. Evaluation Metrics
  9. Privacy
  10. Utility
  11. Conclusion
  12. Limitations and Future Work
  13. Acknowledgments

Figures (3)

  • Figure 1: Left: While voice-only anonymization (using an ASR-TTS pipeline) successfully defeats an audio-only attacker (high EER), it remains vulnerable to a content-only attacker aggazzotti2024can. Right: The content attacker's performance improves as more utterances from the same speaker are aggregated aggazzotti2025asr.
  • Figure 2: EER performance of API-based and local LLM paraphrasing models against a content-based attribution attack. All paraphrasers reduce attribution performance to around chance. Segment-based paraphrasing (GPT5, Gemma3s) preserves privacy more than utterance-based paraphrasing (GPT4o-mini).
  • Figure 3: Left: Synthetic speech detection (SSD) is significantly better than synthetic text detection (STD) on fully anonymized speech/text, especially with only a few utterances. Right: Better STD on original (raw) paraphrases suggests that synthesis/re-transcription removes some machine style.