Perturbed Public Voices (P²V) Overview

• P²V is the study of deliberate and emergent alterations in public speech signals driven by privacy needs, socio-political interventions, and adversarial manipulations.
• It integrates computational modeling, speech processing, and machine learning techniques to analyze sentiment dynamics and detect deepfakes under realistic perturbations.
• Research in P²V enhances privacy-preserving voice anonymization, develops robust adversarial defenses, and supports community-based moderation for resilient public discourse.

Perturbed Public Voices (P$^{2}$V) denote the deliberate or emergent alteration of public speech signals—including text and audio—driven by privacy, security, socio-political intervention, or adversarial manipulation. Research on P$^{2}$V spans computational modeling of collective sentiment in crises, privacy-preserving voice technologies, adversarial countermeasures to synthetic speech abuse, community-centric detection of violence-provoking content, and robust deepfake detection under environmental and algorithmic perturbations. This field synthesizes methods from epidemic modeling, speech signal processing, machine learning, and social computation to capture, modify, protect, and analyze the voice of the public across technological and societal domains.

1. Conceptual Foundations: Modeling and Dynamical Systems

Foundational work models the propagation and escalation of public sentiment and violence by extending epidemic frameworks to multi-agent, multi-state systems. The canonical five-agent ODE model ["From Public Outrage to the Burst of Public Violence: An Epidemic-Like Model" (Nizamani et al., 2013)] distinguishes five classes:

• S (Sensitive): Susceptible individuals
• U (Upset): Those actively propagating outrage
• I (Immune): Exposed individuals rendered resistant, with possible reversion
• V (Violent): Engaged in acts of violence
• R (Relaxed): Recovered and no longer propagating

Core ODEs:

$$\begin{align*} &\frac{dS}{dt} = -(α + β + γ)SU - μSV \ &\frac{dU}{dt} = αSU - ξU + κUI - σU^2 + μSV \ &\frac{dI}{dt} = βSU - κUI \ &\frac{dV}{dt} = γSU - ηV + σU^2 \ &\frac{dR}{dt} = ξU + ηV \end{align*}$$

The model captures two-stage outbreaks—from outrage to violence—quantified by analytic and regression-based thresholds such as the reproductive number $\mathcal{R} = \frac{α}{ξ}$ and empirical peak estimation formulas. This modeling enables scenario analysis and mitigation strategies via rate modulation (contact reduction, accelerated relaxation).

2. Speech and Text Anonymization: Privacy-Preserving Techniques

Technical research on perturbed speech signals centers on anonymization and adversarial perturbation. VoicePrivacy 2022 Challenge (Tomashenko et al., 2022) formalizes the problem as concealing speaker identity while maintaining content and intelligibility. Key approaches include:

• X-vector modification with neural waveform resynthesis (Baselines B1.a, B1.b)
• Acoustic feature anonymization (F0, bottleneck, LPC analysis)
• Randomized formant (McAdams coefficient) shifting (Baseline B2)

Utility and privacy are assessed via Equal Error Rate (EER), Word Error Rate (WER), pitch correlation ($ρ^{F_0}$), and gain of voice distinctiveness ($G_{VD}$).

Selective perturbation frameworks operate at the representation level. ["A Speech Representation Anonymization Framework via Selective Noise Perturbation" (Tran et al., 2022)] injects Laplacian noise into representation dimensions flagged by a Transformer-based privacy-risk saliency estimator, balancing privacy (ASV EER) and utility (ASR WER, emotion classification, intent recognition).

3. Adversarial Defenses and Attacks: Deepfake and Surveillance Robustness

Adversarial perturbation is leveraged both for defending privacy and attacking anonymization. "Stop Bugging Me! Evading Modern-Day Wiretapping Using Adversarial Perturbations" (Mathov et al., 2020) demonstrates real-time UAP (Universal Adversarial Perturbation) methodologies for fooling ASR and topic identification systems in VoIP, employing external hardware to assure threat model resilience.

Mitigation against unauthorized TTS and VC-based voice synthesis is advanced by Pivotal Objective Perturbation (Zhang et al., 28 Oct 2024), which poisons voice training data by directly minimizing generator reconstruction loss during TTS training, ensuring transferability and robustness against augmentation and noise reduction. Objective metrics such as WER and MCD confirm the effectiveness and cross-model applicability of this approach.

Recent multimodal attack models (VoxATtack, (Aloradi et al., 16 Jul 2025)) combine ECAPA-TDNN acoustic speaker embedding with BERT-based textual embeddings, fused via per-sample confidence weighting. Such methods expose fundamental vulnerabilities in current anonymization protocols, revealing speaker identity through preserved text idiolects and robustly attacking diverse systems.

4. Dataset Innovation: Realistic Deepfake Detection Benchmarks

P$^{2}$V also denotes a new standard in dataset creation for evaluation of deepfake and privacy systems. ["Perturbed Public Voices (P$^{2}$V): A Dataset for Robust Audio Deepfake Detection" (Gao et al., 13 Aug 2025)] defines a comprehensive benchmark:

• 247,200 synthetic samples (from 10 cloning pipelines and identity-consistent LLM transcripts)
• 10,240 authentic audio samples from public figures
• 10 types of adversarial/environmental audio perturbations (noise, compression, filtering, time stretching, etc.)
• Cross-domain splits by subject to prevent data leakage

Experiments demonstrate that state-of-the-art deepfake detectors lose 43% performance on P$^{2}$V versus legacy benchmarks, owing to robust voice cloning and realistic noise. P$^{2}$V-trained systems generalize across domains and resist advanced perturbations and synthesis artifacts, establishing a new standard for detection reliability.

5. Sociopolitical and Community Dynamics: Sentiment, Hostility, and Moderation

Hope speech and violence-provoking speech detection represent sociopolitical applications of P$^{2}$V methodologies. Computationally, polyglot word embeddings (FastText SkipGram, (Palakodety et al., 2019)) and curated phrase lexicons facilitate dynamic trend analysis of sentiment in socio-political crises, where spikes in peace-promoting speech occur in response to escalating conflict.

Community-centric annotation and classifier development (Verma et al., 21 Jul 2024) focus on the nuanced detection of violence-inciting online language, revealing that hate speech detection (F1 = 0.89) far exceeds the difficulty of reliably detecting violence-provoking content (F1 = 0.69) due to contextual and semantic subtleties.

Textual perturbation by human agents (e.g., intentional spelling distortions in cyberbullying, political dialogue) is systematized via the CRYPTEXT toolkit (Le et al., 2023), which tracks, normalizes, and generates wild-type text variants, enabling more robust NLP models and effective moderation.

6. Regulatory, Application-Level, and Ethical Considerations

Preset-Voice Matching (PVM, (Platnick et al., 18 Jul 2024)) exemplifies privacy-regulated speech-to-speech translation by substituting arbitrary voice cloning with feature-based preset matching. Hierarchical classifiers extract speaker metadata (gender, emotion), which indexes into a consented preset voice library for TTS output. PVM avoids legal deepfake liabilities, improves naturalness (NISQA metrics), and scales efficiently in multi-speaker environments.

Probabilistic frameworks for analyzing public remarks (Xu et al., 14 Mar 2025) use joint inference (Probabilistic Soft Logic) modeling of meeting structure, speaker roles, and textual transitions. These systems show substantial improvements over text-only classifiers for extracting, categorizing, and analyzing "perturbed" civic voices, supporting scalable, transparent local government accountability.

7. Synthesis and Future Directions

P$^{2}$V research construes public speech and text as fluid, perturbed signals subject to both protective and adversarial interventions. Strategies encompass mathematical modeling, adversarial and representation-level perturbations, multimodal attacks, community annotation, and probabilistic joint reasoning. Robust evaluation under realistic, noisy, and cross-domain conditions is a central focus, with future directions frequently proposed as integrative, privacy-preserving, and socially-aware perturbation mechanisms.

This multi-disciplinary perspective reveals that the voice of the public—whether in civic forums, social media, or private communications—requires both technological and social safeguarding against misuse, deepfake, surveillance, and escalation, while offering new opportunities for democratic transparency and resilient discourse.