Dementia Speech Privacy Beyond English: Where Timing Both Identifies and Diagnoses

Speech is one of the few cheap, non-invasive windows onto neurodegeneration, the family of diseases in which failed mitophagy (the cell's selective recycling of damaged mitochondria) is a recurring cause of neuron loss. Using speech this way requires sharing recordings across hospitals and countries, and a recording is also a biometric identifier. Voice anonymization tools rewrite a recording so the speaker cannot be recognized while the words stay intact. These tools are designed, tuned, and scored almost entirely on English, with privacy and utility judged by an English speaker-verification attacker and an English speech recognizer. This proposal claims that the privacy-utility balance of dementia-screening speech depends on the language, for a specific reason. Speaker timing is the channel attackers use to re-identify people who have been anonymized. It is also the channel that carries many dementia signs, such as pauses, slowed phoneme rate, and hesitations. Languages spread information across timing, pitch, and word choice in different proportions. An anonymizer that looks well balanced in English can therefore leak identity, or erase the diagnosis, in Mandarin or Spanish.

In two sentences: we measure the three-way curve of re-identification risk, transcription quality, and dementia-detection accuracy for existing anonymizers in several languages and under strong attackers. We then test whether controlling the timing channel separately, per language, moves that curve outward. Linguistic publication bias enters as the thing being corrected. The evidence base for "safe to share" clinical speech has the same English-only blind spot that bibliometric reviews of mitochondrial disease already admit to. Mitophagy dysfunction enters only as the clinical motivation. We do not claim speech measures mitophagy.

Why this gap is real

The motivating evidence comes from three places. On the biology side, NR1D1 Inhibition Enhances Autophagy and Mitophagy in Alzheimer's Disease Models links impaired mitophagy to neuronal decline in Alzheimer's models. The research file lists the lack of any non-invasive surrogate for tracking this decline in living patients as an open problem, since current mitophagy assays need tissue. Speech is a candidate surrogate for the downstream cognitive decline, not for the organelle process itself.

On the privacy side, Analysis of Speech Temporal Dynamics in the Context of Speaker Verification and Voice Anonymization shows that speakers can be re-identified from their phoneme pacing even after their voice timbre is masked. You Are What You Say shows the words themselves leak identity. Both results are English-centric.

On the evidence side, the mitochondria-in-myocardial-ischemia bibliometric study A bibliometric and visualized analysis of research on mitochondria in myocardial ischemia from 2015 to 2024 finds China produced over 47% of indexed output. It also concedes that excluding non-English work is a major limitation. The cognitive-impairment bibliometrics of Bo Zou et al. make the same admission. Much of the clinical population these tools would serve does not speak English, and neither the literature nor the speech tools look there.

The closest existing work pulls in the opposite direction from ours. ClaritySpeech: Dementia Obfuscation in Speech treats dementia as the private attribute. It transcribes with Whisper, paraphrases the text to remove dementia markers, and re-synthesizes with a zero-shot text-to-speech model that flattens irregular timing. That is the right design when the goal is to hide a diagnosis. It destroys exactly what a screening study needs to keep.

MUSA makes anonymization multilingual through serial disentanglement. It does not evaluate diagnostic utility or timing-based attackers. Our proposal treats dementia signs as the utility to preserve and identity as the thing to remove. The evaluation is carried out per language. The timing channel is controlled explicitly rather than handed intact to a vocoder, the neural network that turns the extracted features back into audio.

Core Claim

Timing features carry both identity and dementia signs, and this overlap is language-dependent. A pipeline that passes timing through unchanged keeps diagnostic accuracy but stays vulnerable to timing-aware attackers. A pipeline that normalizes timing is safer but loses the diagnosis. Where the best compromise lies differs between a stress-timed language like English, a syllable-timed one like Spanish, and a tonal one like Mandarin. In Mandarin, pitch also carries word meaning, so pitch cannot be altered freely without garbling the content.

The testable prediction has two parts. First, the ranking of existing anonymizers on a joint privacy-and-diagnosis score will change across languages. Second, a per-language timing transform that keeps pause and rate statistics at the group level while scrambling each speaker's individual rhythm will dominate both pass-through and full normalization in at least the non-English languages.

Research Program

The first component is a measurement protocol. We extend the VoicePrivacy evaluation described in The Third VoicePrivacy Challenge with an extra utility axis: dementia detection. This is a classifier trained on original recordings, following the challenge's practice for emotion. We also add a second classifier retrained on anonymized recordings. Privacy is measured as equal error rate, the operating point where a speaker-verification attacker's false accepts equal its false rejects; 50% means the attacker is guessing. We report it under three attacker strengths.

The first attacker knows nothing about the anonymization. The second is an informed attacker retrained on anonymized audio, following Evaluating Voice Conversion-Based Privacy Protection against Informed Attackers. The third combines the dual-stream design of DAST, the feature attacks of Attacking Voice Anonymization Systems with Augmented Feature and Speaker Identity Difference, and an explicit phoneme-duration stream that follows the temporal-dynamics analysis. Content is measured by word error rate for English and Spanish and character error rate for Mandarin.

The second component is the intervention. We start from a disentangle-and-resynthesize pipeline of the kind introduced by Speaker Anonymization Using X-vector and Neural Waveform Models, using self-supervised content features as in Reprogramming Self-supervised Learning-based Speech Representations. We then separate out the duration and pause track as its own input to resynthesis.

The transform on that track keeps a speaker's coarse statistics, such as mean pause length, pause rate, and speech rate. These are the quantities dementia classifiers mostly rely on. It resamples fine-grained phoneme durations from a language-specific population model, which removes the idiosyncratic rhythm that verification attackers exploit. For Mandarin, pitch is split into a tone-carrying lexical contour, which is kept, and a speaker-level register, which is shifted. The only per-language tuning is fitting the duration population model on healthy-speaker data from that language.

The third component is a small meta-scientific audit that ties the work back to publication bias. We tabulate the languages in which published anonymization and dementia-speech systems in the file were evaluated. We then ask how often an "acceptable trade-off" conclusion drawn from English results survives in our other languages. This turns the language-bias concern into a measured failure rate rather than a caveat.

Baselines

The naive reference point is releasing the original audio with no anonymization. It sets the ceiling for dementia detection and content accuracy and the floor for privacy.

The standard pipeline baseline is the x-vector and neural-waveform approach of Fang et al., which passes duration through untouched. The file reports that early baselines of this type raised ASR word error rate by 19% to 67% (Introducing the VoicePrivacy Initiative). It also reports that near-50% equal error rate systems in the 2024 challenge struggled to preserve emotions such as sadness. We report the file's figures only in these qualitative terms and re-measure everything ourselves.

The strongest multilingual competitor is MUSA, including its zero-embedding identity setting. It is the only system in the file built for multiple languages. We run its released configuration on all three languages. The strongest competitor on the diagnosis-related axis is ClaritySpeech, run as designed. It should score high on privacy and near chance on dementia detection, which marks the opposite end of the curve. No reported result was found on Lacuna for any of these systems on dementia-detection utility or on non-English dementia speech. TVTSyn, which models timbre over time, serves as a check that our gains come from timing and not from timbre modeling.

The key ablation compares three otherwise identical versions of our pipeline: duration passed through, duration fully normalized to a population average, and our statistics-preserving resampling. This isolates the proposed ingredient. A second ablation applies an English-fitted duration model to Mandarin and Spanish, to test whether per-language fitting matters.

Every system gets the same input audio, the same attackers and classifiers, and the same number of hyperparameter configurations tuned on a held-out development split. We count it as beating a baseline only if our system improves dementia-detection score at equal or higher informed-attacker equal error rate, or improves that error rate at equal detection score, with non-overlapping speaker-level bootstrap intervals, in at least two of three languages.

Why not just…?

Why not just run MUSA per language and report the numbers? That is our strongest baseline and part of the protocol, so it is not an alternative so much as the first experiment. MUSA was not evaluated against timing-aware attackers or for diagnostic utility. If it already sits on the best curve in every language, our intervention is unnecessary, and we would report that plainly.

Why not just share transcripts or handcrafted diagnostic features instead of audio? This is a real threat. Sharing only pause statistics and a small set of lexical measures may give dementia accuracy close to full audio with little identity exposure. We include it as a baseline, attacked with the lexical approach of You Are What You Say and a duration-statistics attacker. If it matches our pipeline, the honest conclusion is that waveform release is not worth its risk for screening. The remaining case for audio would then be reuse for tasks nobody specified in advance.

Why not just fix fairness in the evaluation, testing English tools on translated or synthetic non-English speech? Synthetic speech has no real pathological timing, and zero-shot synthesis normalizes irregular timing, which is the very effect ClaritySpeech exploits. Such data would hide the effect we are trying to measure. The comparison needs real recordings of people with dementia in each language.

Proposed Experiments

We run the English stage first. It uses the VoicePrivacy protocol for privacy and content, plus the dementia speech corpora of the kind used by ClaritySpeech for the diagnostic axis. The non-English stage needs Mandarin and Spanish recordings with clinical labels. This is the main risk: such corpora are scarcer and more tightly access-controlled, which is itself a symptom of the bias under study. If labeled Mandarin data cannot be obtained, we fall back to healthy Mandarin speech for the privacy and content axes and a small partner-collected clinical set for detection, and we report the reduced statistical power.

The decisive checks are three. First, rank reversal: rank correlation below 0.5 on the joint score between English and at least one other language would confirm that English-only evaluation misleads. Second, the timing ablation: if our resampling does not beat both pass-through and full normalization, the core claim fails. Third, an attacker-strength check: our system's privacy advantage should grow, not shrink, when the attacker gains the explicit duration stream, since that stream is the one we alter. If the privacy advantage disappears under that attacker, per-speaker rhythm is still leaking through the other channels, and the claim that timing is the shared channel is wrong.