Reading the Gut from the Voice: A Prosody Codec as a Non-Invasive Readout of Microbiome Interventions

The bacteria in our intestines influence the brain. This is not correlation hand-waving: when fecal material from human depression patients is transplanted into microbe-free rodents, the recipients develop depressive-like behavior (Teng et al.), and fecal microbiota transplantation (FMT) — literally transferring a donor's entire intestinal microbial community into a recipient — suppresses neuroinflammation after traumatic brain injury in mice (Fecal microbiota transplantation inhibited neuroinflammation of traumatic brain injury in mice via regulating the gut–brain axis). Meanwhile, changes in the brain reliably leave fingerprints in how a person speaks: pause patterns, pacing, and pitch contours carry enough signal to detect Alzheimer's dementia from spontaneous speech (Alzheimer’s Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge). No study has ever connected these two facts — nobody has measured whether a gut-microbiome intervention shifts human vocal production, or whether speech can serve as a non-invasive, at-home readout of such interventions.

The bridge is a surprising piece of speech technology. Zero-shot voice cloning systems synthesize a target speaker's voice from a few seconds of audio, and the best of them internally split a voice into separable factors: who is speaking (timbre), what is said (content), and how it is delivered — pacing, pauses, pitch movement, collectively prosody. Systems like PACE decouple prosody from timbre by minimizing their mutual information (Prosody-Adaptable Audio Codecs for Zero-Shot Voice Conversion via In-Context Learning). My proposal is to repurpose that prosody factor as a measurement instrument: a per-utterance "physiological state" signal that is invariant to speaker identity but sensitive to the neuro-muscular and cognitive state that the gut–brain axis can alter. One caveat shapes the whole design: ClaritySpeech: Dementia Obfuscation in Speech showed that a vanilla zero-shot TTS system actively smooths atypical prosody toward fluent-speech norms — the pathology signal is exactly what it deletes — so the naive version of this idea fails, and the fix is deliberate disentanglement.

Core Claim

A frozen, prosody-disentangled voice codec can be used as a longitudinal physiological readout of microbiome interventions in humans. Concretely: the codec's prosody-state vector, extracted from standardized weekly voice recordings of patients undergoing FMT, will (a) be invariant to speaker identity by construction, (b) shift measurably after the intervention, (c) predict the patient's post-intervention gut-microbial community embedding better than classical acoustic feature sets, and (d) resist shortcut confounds such as age and BMI when the linking model is constrained to preserve causal structure, following CARL: Preserving Causal Structure in Representation Learning. A secondary instrument falls out for free: the distance between a speaker's original prosody-state and the prosody-state an off-the-shelf zero-shot TTS system "corrects" it toward — the normalization effect ClaritySpeech demonstrated — serves as a label-free index of vocal atypicality, turning the known bug of zero-shot TTS into a measurement.

The causal motivation comes from rodent FMT work (Teng et al.; Fecal microbiota transplantation inhibited neuroinflammation of traumatic brain injury in mice via regulating the gut–brain axis), but the measurement target is deliberately human. Human speech codecs are built on human phonetics and text alignment; rodent ultrasonic vocalizations are out of their reach, so all empirical claims here are restricted to human interventional cohorts. This is what separates the proposal from the adjacent literature: Teng et al. transfer psychiatric phenotypes but never record vocal behavior; ADReSS and ClaritySpeech measure pathology in speech but never tie it to a causal biological intervention; CARL preserves causal structure in multi-omic data but has never touched acoustics. Nothing on Lacuna connects an in-vivo microbiome intervention to zero-shot voice modeling — the file reports zero studies at this intersection.

Research Program

The work has three components. First, the frontend: we adopt the PACE architecture (Prosody-Adaptable Audio Codecs for Zero-Shot Voice Conversion via In-Context Learning) — an audio codec whose encoder decouples prosodic contours from speaker timbre via mutual-information minimization — train it on LibriTTS-clean-100, verify it reproduces the published operating point (speaker-verification similarity 0.9078 versus 0.8429 for VALL-E X, prosody-matching $f_0$ distance 2.6988 versus 3.1029, word error rate 0.1010, mean opinion score 4.36 versus 4.39 for ground truth), and then freeze it. Freezing matters: our cohorts will have tens of subjects, and the Waypoint scaling study showed random forests beat deep models below roughly 1,000 samples, with deep models only taking the lead above about 10,000 (LLMs Can Learn the Language of the Microbiome). All deep modeling happens at pretraining scale; nothing is trained on the cohort except small probe heads.

Second, the cohort: roughly 60–80 irritable bowel syndrome patients undergoing clinically indicated FMT, following the intervention protocols of the existing IBS FMT literature (Long-Term Follow-Up Results of Fecal Microbiota Transplantation for Irritable Bowel Syndrome: A Single-Center, Retrospective Study; Poster Session II A324 FECAL MICROBIOTA TRANSPLANTATION FOR SYMPTOM IMPROVEMENT IN PATIENTS WITH IRRITABLE BOWEL SYNDROME), plus their healthy donors. Each participant gives a standardized read-aloud voice recording and a stool sample weekly for four weeks before and eight weeks after transplantation, alongside serum inflammatory markers and cognitive screening. The design is deep-phenotyping in the spirit of the Human Phenotype Project, which already combines gut metagenomics with multiple clinical modalities and shows diet reshapes the microbiome longitudinally (Diet shapes the gut microbiome: cross-sectional and longitudinal insights from the Human Phenotype Project) — we add the one modality it lacks: repeated structured voice.

Third, the analysis. Stool samples are encoded with self-supervised metagenomic embeddings, which at low sample sizes beat raw abundances on clinical traits (age prediction Pearson $r = 0.14$ for scVI embeddings versus $0.06$ for raw abundances at $n = 100$; drug-usage classification PR-AUC $0.81$ versus $0.73$) and improve cross-cohort generalization in 82% of comparisons (Self-supervised learning enables robust microbiome predictions in data-limited and cross-cohort settings). We then test four things: within-subject pre/post classification from the prosody-state vector; prediction of the post-FMT microbial embedding from the prosody-state; donor–recipient convergence (does the recipient's prosody-state move toward the donor's after transplantation — the human analog of phenotype transfer); and a mediation analysis under CARL-style conditional-independence and Markov-boundary constraints, so that any prosody–microbiome link is checked against shortcuts through age, BMI, or baseline severity, the failure mode CARL documents for unconstrained representation learning (CARL: Preserving Causal Structure in Representation Learning). Finally, because the codec is generative, we can synthesize identity-fixed audio that varies only the prosody-state and ask blinded clinicians to rate whether post-FMT samples sound different from pre-FMT samples — a validation classical feature sets cannot produce.

Baselines

The naive reference point is chance performance on within-subject pre/post discrimination (50% accuracy, AUC 0.5), which anchors how much longitudinal signal exists at all. The strongest classical competitor is the standard clinical-acoustics pipeline: openSMILE-style feature sets (ComParE or eGeMAPS) fed to a linear discriminant or random forest, exactly the ADReSS baseline configuration, which achieved 0.625 test-set accuracy for dementia classification and MMSE regression RMSE of 6.14 with decision trees on MRCG features (Alzheimer’s Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge). This is the fiercest baseline because our cohorts are small and the Waypoint scaling result favors shallow models in exactly this regime (LLMs Can Learn the Language of the Microbiome); we run it on the identical recordings, folds, and tuning budget, with per-speaker z-normalization as its best-practice variant for removing identity. The strongest deep baseline is a global speaker embedding from a non-disentangled zero-shot TTS stack (the VALL-E X configuration, speaker-verification similarity 0.8429 and $f_0$ distance 3.1029 on the same benchmark where PACE reaches 0.9078 and 2.6988; Prosody-Adaptable Audio Codecs for Zero-Shot Voice Conversion via In-Context Learning), because it has the same capacity and training data as our frontend but lacks only the prosody/timbre split. The key ablation is our own architecture with the mutual-information disentanglement loss removed: same codec, same data, same probes — this isolates disentanglement as the active ingredient. On the microbiome side, raw taxonomic abundances serve as the embedding baseline (Self-supervised learning enables robust microbiome predictions in data-limited and cross-cohort settings). We count the proposal as successful only if the prosody-state beats speaker-normalized classical features on within-subject change detection (target: AUC margin of at least 0.05 with confidence intervals excluding zero), beats the non-disentangled embedding by a similar margin, and shows a significant donor-convergence effect under permutation testing. For completeness we note that per-task numeric scores on the Compass microbiome benchmark are not available — no reported result found on Lacuna for Compass sub-task scores — so we will not use Compass for microbiome-embedding selection.

Why not just…?

Why not just extract classical acoustic features and train a shallow classifier? This is the most serious alternative, and the honest answer is that it might work: classical acoustics already detect dementia at 0.625 accuracy (Alzheimer’s Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge), and shallow models win in small-sample regimes (LLMs Can Learn the Language of the Microbiome). But classical feature sets entangle timbre with prosody, so between-speaker variation dominates the between-time variation we care about; the classical fix, per-speaker normalization, requires enough baseline recordings per speaker and discards any information shared across speakers. The disentangled codec performs that separation by construction and additionally enables counterfactual synthesis. Crucially, we are not arguing around the alternative: speaker-normalized classical features are our primary baseline, and if they match the prosody-state's AUC on within-subject change detection, the experiment settles it against us and we will report that.

Why not just measure blood biomarkers and cognitive scores instead of speech? Serum inflammatory markers and standard exams are cheaper, better validated, and already collected in our cohort — as covariates, not competitors. The case for speech is sampling density and continuity: a weekly one-minute recording is essentially free for the patient at home, while blood draws are not. The claim we must defend empirically is incremental: a nested-model comparison testing whether prosody-state adds predictive power for the post-FMT microbial embedding beyond biomarkers and cognitive scores. If it adds nothing, the readout is redundant and only the counterfactual-audio capability remains as a contribution.

Why not just use an off-the-shelf zero-shot TTS system on the recordings? Because it destroys the signal. ClaritySpeech: Dementia Obfuscation in Speech demonstrated that a zero-shot TTS system treats atypical prosody as noise: rerouting dementia speech through XTTSv2 dropped adversarial dementia-detection F1 from 0.72 to 0.59 while cutting word error rate from 0.73 to 0.08 and retaining only 50% speaker similarity. Using the vanilla system as our analyzer would erase exactly the physiological signal we want to track. Only a codec with explicit prosody disentanglement, in the PACE or negated-representation style (Prosody-Adaptable Audio Codecs for Zero-Shot Voice Conversion via In-Context Learning; Enhancing Zero-Shot Multi-Speaker TTS with Negated Speaker Representations), can hold timbre fixed while preserving the state factor. And we turn the erasure itself into a secondary instrument: the magnitude of the normalization pull, measured as the distance between original and re-synthesized prosody-states, is a label-free atypicality index we validate against clinical ratings.

Decisive Checks

Three results would make or break the proposal. First, the within-subject AUC comparison: prosody-state versus speaker-normalized classical features, pre-registered, on held-out subjects — a win for classical features falsifies the central claim. Second, the donor-convergence permutation test: if recipients' post-FMT prosody-states are no closer to their donors' than permutation allows, the phenotype-transfer story does not extend to human vocal production, and the proposal collapses into a weaker correlation study. Third, the shortcut challenge: re-running the prosody-to-microbiome prediction with age, BMI, and baseline severity forcibly balanced or held out, under CARL-style constraints (CARL: Preserving Causal Structure in Representation Learning) — if the link survives only in unbalanced data, we have rediscovered a confound, not a gut–voice axis. Each check is designed so that a negative answer is informative rather than fatal to the field: it would still tell us whether the microbiome's reach, so clearly demonstrated in rodent brains, is audible in the human voice.