Content Anonymization for Privacy in Long-form Audio

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Existing voice anonymization techniques, successful for short utterances in benchmarks like the VoicePrivacy Challenge, are insufficient for long-form audio such as interviews or phone calls. This is because an attacker can exploit linguistic content—vocabulary, syntax, and turns of phrase—to re-identify speakers, even with disguised voices. This research proposes novel content anonymization approaches that perform contextual rewriting of Automatic Speech Recognition (ASR) transcripts within an ASR-Text-to-Speech (TTS) pipeline. The method uses LLM-based paraphrasing, specifically segment-based paraphrasing with models like Gemma-3-4B or GPT-5, to eliminate speaker-specific style while preserving meaning. Experiments on the Fisher Speech Corpus demonstrate that content-based attacks are effective against voice-only anonymization, but the proposed joint voice and content anonymization achieves an Equal Error Rate (EER) exceeding 50%, effectively mitigating this risk. The anonymized speech also shows improved naturalness with a UTMOS score of 3.14 compared to 2.09 for original audio.

Key takeaway

For MLOps Engineers deploying voice anonymization systems for long-form audio, you must integrate content anonymization to prevent re-identification via linguistic style. Implement contextual, LLM-based paraphrasing within your ASR-TTS pipeline. Prioritize segment-based approaches for robust privacy, ensuring defense against content-based attacks and achieving an EER exceeding 50%.

Key insights

Linguistic content in long-form audio is a privacy risk, requiring content anonymization beyond voice disguise.

Principles

• Long-form audio content is a biometric side-channel.
• Voice-only anonymization is insufficient for long-form audio.
• Contextual paraphrasing mitigates content-based re-identification.

Method

An ASR-TTS pipeline rewrites transcripts using LLM-based contextual paraphrasing on a sliding window of multiple utterances to eliminate speaker-specific style.

In practice

• Implement LLM-based paraphrasing in ASR-TTS pipelines.
• Use segment-based paraphrasing for better privacy.
• Consider open-weight models like Gemma-3-4B for on-device privacy.

Topics

• Voice Anonymization
• Content Anonymization
• Long-form Audio Privacy
• ASR-TTS Pipelines
• LLM Paraphrasing
• Speaker Re-identification

Code references

• openai/whisper

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, MLOps Engineer, AI Security Engineer

Related on AIssential

• How to Leverage Synthetic Speech for LLM-Based ASR Systems? — RIR-augmented synthetic speech, combined with layer-wise pooling, significantly reduces real data needs for LLM-based ASR.
• US scrambles to stop Internet users re-creating dead pilots’ voices — Spectrograms combined with AI tools enable unauthorized audio reconstruction, challenging privacy laws.
• Introducing VibeVoice ASR: Longform, Structured Speech Recognition At Scale — VibeVoice ASR unifies long-form speech transcription, diarization, and timestamping into a single, context-preserving model.
• WordVoice: Explicit and Decoupled Multi-Dimensional Word-Level Control for LLM-Based TTS — WordVoice enables precise, decoupled word-level acoustic control in LLM-based TTS through explicit planning and a new annotated dataset.
• Large Audio Language Models for Spoofing-Aware Speaker Verification — LALMs can achieve competitive Spoofing-Aware Speaker Verification with adaptation, offering auditable rationales.
• HW-TSC’s Submissions to the IWSLT 2026 Offline Speech Translation Task — A two-pass transcription and context-aware merging strategy improves long-form speech translation quality.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.