Two research communities currently work past each other. Speech researchers can measure the integrity of motor speech — how intelligibly a person with Parkinson's disease hypokinetic dysarthria speaks — with ever finer resolution, from phoneme-level probability smearing inside speech recognizers to disentangled latent embeddings that isolate what was said from who said it. Gut–brain researchers can do something speech researchers cannot: establish causality. By transplanting a complete fecal microbial community from a human donor into a microbiologically cleared animal (fecal microbiota transplantation, FMT), they have shown that a diseased microbiome alone is sufficient to reproduce central neural pathology — depleting amino acids in the prefrontal cortex and downregulating glutamate transporters in rats, or inducing pro-thrombotic vascular pathology in mice. Nobody has ever connected the two: no study transplants microbiota and asks what happens to the neuromuscular control of speech, and no study uses speech measurement as a functional readout of microbiome interventions in humans.
This proposal does both, in two coupled arms. The core claim: fine-grained, channel-normalized speech intelligibility metrics — the kind already validated against human listener panels — can serve (a) as a causal probe of whether dysbiotic gut microbiota from patients with Parkinsonian dysarthria transmit articulatory motor degradation to recipient animals, and (b) as a non-invasive, high-frequency longitudinal biomarker of patient response and durability in clinical FMT trials, where the current evaluation bottleneck is severe. The clinical urgency of (b) is concrete: in a randomized trial of oral FMT capsules for hypertension, the blood-pressure effect measured −4.34 mmHg at Day 7 ($p = 0.024$) but had completely vanished by Day 30 ($p = 0.62$), and detecting that collapse required repeated blood draws, mass spectrometry, and metagenomic sequencing of 748 stool samples Fecal microbiota transplantation for hypertension. A weekly five-minute recorded sentence set that tracks the same trajectory would change how such trials are monitored.
Core Claim
The claim has a biological half and a measurement half, and the proposal stands or falls on both.
Biologically: motor speech deficits in Parkinson's disease are treated in the speech literature as a purely intracranial, dopamine-depleted basal-ganglia phenomenon A Multitask Learning Approach to Assess the Dysarthria Severity in Patients with Parkinson's Disease. But gut–brain work has shown that microbial metabolites actively modulate neuroinflammation, blood–brain barrier permeability, and synaptic transporter expression in the central nervous system Microbiota alterations leading to amino acid deficiency contribute to depression in children and adolescents; Cold environment regulates ischemic stroke through modulation of gut microbiota. The untested hypothesis is that a measurable fraction of bulbar (speech-muscle) motor degradation is microbiome-mediated — testable by transplanting stool from dysarthric PD donors versus matched controls into antibiotic-depleted rats and measuring bulbar endpoints directly.
On measurement: speech is uniquely positioned as a biomarker because it is produced daily, non-invasively, and at millisecond motor resolution. The critical methodological constraint comes from a warning already in the literature: standard dysarthric speech corpora are environmentally confounded — in UA-Speech, healthy controls were recorded at a mean SNR of +3.7 dB while patients were recorded at −7.6 dB, and classifiers achieved 96–100% accuracy on non-speech silence segments alone, proving they learned the recording room, not the pathology On using the UA-Speech and TORGO databases to validate automatic dysarthric speech classification approaches. Any longitudinal speech biomarker for an FMT trial must therefore be built around within-speaker comparison under a channel-normalized recording protocol, or it will measure microphone placement, not neuromotor recovery.
Research Program
Arm 1 (preclinical, causality). Recruit PD patients with mild-to-moderate hypokinetic dysarthria and matched healthy controls; record each donor's speech under the controlled protocol below and score it with the disentangled-embedding and transcription metrics. Transplant donor stool into antibiotic-depleted recipient rats following the design that established causality for depression — human donors, rigorously phenotyped, recipients cleared with broad-spectrum antibiotic cocktails Microbiota alterations leading to amino acid deficiency contribute to depression in children and adolescents. Rodents cannot produce human speech, so the bulbar-relevant endpoints are the mechanistic ones that paper's machinery makes tractable: amino acid concentrations and EAAT2/EAAT3 glutamate transporter expression in brainstem motor nuclei and motor cortex, laryngeal/tongue muscle histology, neuroinflammatory cytokines, and acoustic analysis of rat ultrasonic vocalization timing as a coarse cross-species motor-speech proxy. The decisive readout is the dose–response relationship between donor speech severity (continuous, from the metrics below) and recipient bulbar molecular markers — if donor dysarthria severity predicts recipient deficits, the microbiome carries articulatory motor information.
Arm 2 (clinical, biomarker). A 90-day longitudinal speech-monitoring study in PD patients with dysarthria undergoing FMT versus placebo (FMT is already administered to PD patients for gastrointestinal comorbidity, making an add-on speech-monitoring design ethically straightforward). Weekly recordings in a fixed room with a fixed, calibrated microphone; every session logs a silence-segment noise floor so channel drift is quantified and correctable. Each session yields a battery: closed-set rhyme-style minimal-pair discrimination that strips away semantic guessing, which otherwise lets listeners and recognizers compensate for degraded acoustics with context Behavioral Paradigms for Measuring Speech Intelligibility; crowdsourced Visual Analogue Scale ratings, which are the most reliable human modality (interrater reliability D-coefficient = 0.90, versus 0.47 for transcription word accuracy Speech Intelligibility of Dysarthric Speech); and the automated metrics described under Baselines. The design explicitly interrogates which intelligibility construct moves: if FMT stabilizes neuromotor control, the gain should register in phonetic-level metrics; if it eases vocal effort, it should register in clinician-perceived quality — a distinction the measurement literature says is routinely conflated (experts rate patients roughly 30% below transcription scores because they penalize voice quality Comparison of 5 methods for the evaluation of intelligibility in mild to moderate French dysarthric speech). The 90-day window is chosen deliberately to straddle the Day-7-to-Day-30 durability failure documented in the hypertension trial Fecal microbiota transplantation for hypertension.
A prerequisite study precedes both arms: test–retest characterization of every metric in untreated PD speakers over 4 weekly sessions under the controlled protocol. A metric with poor within-speaker test–retest reliability cannot detect intervention effects, whatever its cross-sectional correlation with human scores. Only metrics with acceptable minimal detectable change proceed to Arm 2.
Baselines
The naive reference point is off-the-shelf automatic transcription: Whisper run on each weekly recording, reporting word error rate and character error rate, with no adaptation, no tuning, and the same audio given to every method. This is what a clinical team would do with zero methodological investment, and it must be beaten to justify anything more complex. The literature already flags its weaknesses concretely — Whisper trails human listeners by 7.9 dB to 8.6 dB in speech reception threshold under reverberation, and degrades 4.9 dB from English to German Hearing deficits of transformer-based ASR for anechoic and spatial signals — but weakness on cross-sectional benchmarks is not the same as weakness on within-speaker change detection, which is why it stays in as an honest competitor. Beating it means showing reliably smaller minimal detectable change and earlier, blinded detection of the FMT-versus-placebo divergence.
The strongest existing method is the disentangled-latent intelligibility index of Disentangled Latent Speech Representation for Automatic Pathological Intelligibility Assessment: a SpeechSplit-style autoencoder trained on healthy multi-speaker audio separates what was said (the phonetic content code $Z_c$) from pitch, rhythm, and speaker identity; a patient's $Z_c$ is aligned to healthy reference speech by dynamic time warping, and the squared difference serves as an intelligibility score. This is the paper's strongest competitor because it reports the best correlation with human intelligibility ratings of any method in the file — Pearson $R = -0.89$ (±0.02) across 26 UA-Speech speakers, and it stays strong ($R = -0.81$ ± 0.05) with only 20 utterances per speaker, meaning it is data-efficient enough for weekly clinical use. It will be run exactly as published: retrained on the same healthy-speech data, applied to the identical recordings given to every other method, with the same feature budget and no extra tuning. It is a cross-sectional design, however — built to rank speakers against references, not to track one speaker over time — so beating it here means demonstrating superior test–retest reliability in stable speakers and smaller minimal detectable change in the longitudinal setting, not a higher correlation with human scores. If it matches our battery on those longitudinal criteria, the simpler published method should be adopted and we will say so.
A second established competitor is the composite regression from the French five-method comparison — a regularized linear model combining acoustic features (jitter, shimmer, fundamental frequency) with DNN-HMM speech-recognizer outputs, which correlated $r = 0.70$ to $0.74$ with all three human evaluation modalities Comparison of 5 methods for the evaluation of intelligibility in mild to moderate French dysarthric speech. It represents the "cheap classical features plus ASR" middle ground and will be trained with the same cross-validation protocol and feature budget on our pilot data.
Finally, an ablation isolates the proposal's key methodological ingredient: the channel-normalized, fixed-environment recording protocol combined with within-speaker scoring. The ablation arm processes the same patients' recordings without channel normalization — or with deliberately introduced channel perturbations of a few dB SNR, calibrated to the 11 dB environmental discrepancy documented in UA-Speech On using the UA-Speech and TORGO databases. If the metrics report spurious intelligibility shifts from channel perturbation alone of the same magnitude as any FMT effect we hope to detect, the protocol ingredient is load-bearing; if the metrics are robust to it, the elaborate protocol can be relaxed.
Why not just…?
Why not just use the standard clinical motor scale (MDS-UPDRS-III) instead of speech metrics? This is the simplest alternative and a real one — the speech item of the clinical motor exam is already collected in every PD trial, and validated vision-based motor severity estimation exists for gait PD-Motor-Severity-Estimation. The counter-evidence is twofold. Clinician ratings carry systematic bias: experts scored Friedreich Ataxia speakers roughly 30% below face-to-face transcription scores because they fold vocal effort and voice quality into "intelligibility" Comparison of 5 methods. And coarse clinical targets are measurably less informative than fine articulatory dynamics: a multitask CNN predicting eleven clinical dimensions from 160-ms time–frequency windows of vocal fold onset/offset transitions raised average unweighted recall from 43.9% (single-task) to 49.4%, and PD-versus-control classification from 79% to 89% on offset transitions A Multitask Learning Approach — bulbar motor timing carries signal the macroscopic scale misses. Still, MDS-UPDRS-III remains in our design as a secondary endpoint precisely so the head-to-head comparison is empirical rather than asserted; if the clinical speech item detects the FMT effect as early and as reliably as the battery, the simpler instrument wins.
Why not just compute word error rate with off-the-shelf ASR? Because both error sources documented in the literature corrupt it exactly where we need it. First, semantic guessing: an open-set recognizer (or human) reconstructs words from sentence context, masking exactly the bottom-up phonetic degradation that a neuromotor intervention would change; closed-set rhyme-style tests exist to remove this Behavioral Paradigms for Measuring Speech Intelligibility. Second, model fragility: Whisper's 7.9–8.6 dB gap to human listeners under reverberation and its 4.9 dB language bias Hearing deficits of transformer-based ASR mean its error rate is dominated by model pathologies, not speaker pathologies, and any longitudinal shift could be a recognizer artifact. That said, this alternative is cheap enough that we run it as the naive baseline throughout; the pilot is explicitly powered to answer whether plain WER suffices, and if its test–retest reliability matches the battery's, we report that the elaborate metrics were unnecessary.
Why not just supplement the implicated metabolites (lysine, butyrate) instead of doing FMT at all? If the mechanism is metabolite depletion, oral supplementation is simpler, safer, and more controllable than transplanting an entire unstandardized microbial community — and for the therapeutic goal it is a genuine long-term rival. What the evidence rules out is metabolite supplementation as a substitute for the causal question Arm 1 asks: in the depression cohort, the microbiome perturbation downregulated 297 of 360 circulating metabolites and dysregulated 13 of 20 standard amino acids — a system-wide collapse in which single-metabolite rescue (lysine restriction alone induced the phenotype, but only one branch of it) cannot establish whether the community is the causal unit Microbiota alterations leading to amino acid deficiency; likewise, cold-driven pathology proceeded simultaneously through barrier-tight-junction loss, LPS endotoxemia, and TMAO elevation, axes a single supplement cannot manipulate together Cold environment regulates ischemic stroke. FMT is the only vehicle that treats the whole ecological community as one independent variable. For therapy, a metabolite arm is a legitimate follow-up that our design does not preclude — Arm 1's mediator analysis (which plasma metabolites statistically mediate any donor-to-recipient effect) is exactly what would nominate candidates for it.
Proposed Experiments and Decisive Checks
The pilot establishes measurement validity on existing benchmarks before any FMT occurs. The full battery — Whisper WER, the SpeechSplit $Z_c$ dynamic-time-warping index, the composite acoustic-ASR regression, and an ASR-posterior smearing score in the spirit of $LCP(2.0)$, which achieved RMSE 8.42 against human intelligibility scores versus 9.97 for STOI and 12.10 for eSTOI Predicting Intelligibility of Enhanced Speech Using Posteriors Derived from DNN-Based ASR System — is first evaluated cross-sectionally on PD speech from the PC-GITA corpus (50 PD patients, 50 matched controls, scored on the modified Frenchay Dysarthria Assessment), where the published reference point for PD-versus-control discrimination is 89% UAR on offset transitions with the multitask CNN A Multitask Learning Approach; we report our battery against that number. The $Z_c$ index is additionally re-validated on UA-Speech against its published $R = -0.89$ Disentangled Latent Speech Representation, with the environmental-confound analysis of that corpus On using the UA-Speech and TORGO databases used as the calibration set for the channel-perturbation ablation. Human-score anchors come from the French dysarthria cohort protocol (face-to-face transcription, crowdsourced multi-judge panels, expert ratings) with its published composite correlation of $r = 0.70$–$0.74$ Comparison of 5 methods and from VAS reliability of D = 0.90 Speech Intelligibility of Dysarthric Speech. No reported result was found on Lacuna for test–retest reliability of any of these metrics in within-speaker longitudinal designs — that is precisely the pilot's contribution.
Arm 1's decisive check is the donor-severity/recipient-deficit correlation with mediator analysis, pre-registered: a significant association between donor speech scores and recipient bulbar molecular endpoints, surviving adjustment for donor clinical covariates, plus identification of mediating metabolite pathways in the depression-FMT paper's analytical style Microbiota alterations leading to amino acid deficiency. A null result — recipient endpoints unaffected by donor dysarthria severity — would itself be informative, constraining gut–brain causality claims in PD and redirecting the program to Arm 2 as purely a biomarker play.
Arm 2's decisive checks are three. First, the placebo-controlled divergence: blinded analysts must recover the FMT-versus-placebo assignment from speech trajectories alone above chance, at or before the timepoint where conventional clinical endpoints register change. Second, the durability trajectory: speech metrics must detect the transient-then-vanishing response pattern documented for blood pressure in the hypertension trial Fecal microbiota transplantation for hypertension — if that pattern recurs in PD, weekly speech monitoring catches it without a single blood draw; if no FMT effect on speech exists, the metrics must show flat trajectories indistinguishable from placebo, demonstrating they do not manufacture effects. Third, construct dissociation: which metric family moves (phonetic-level, transcription-level, or quality-perception-level) versus which stays flat, delivering the first empirical answer to how biological interventions map onto the distinct intelligibility constructs the measurement literature keeps conflating Speech Intelligibility of Dysarthric Speech; Comparison of 5 methods. Together, the two arms either establish speech as a causal readout of the gut–brain axis in motor neurodegeneration, or they rule it out cleanly — and both outcomes are publishable, falsifiable science.
References
- Fecal microbiota transplantation for hypertension: an exploratory, multicenter, randomized, blinded, placebo-controlled trial
- A Multitask Learning Approach to Assess the Dysarthria Severity in Patients with Parkinson's Disease
- Microbiota alterations leading to amino acid deficiency contribute to depression in children and adolescents
- Cold environment regulates ischemic stroke through modulation of gut microbiota
- On Using the UA-Speech and Torgo Databases to Validate Automatic Dysarthric Speech Classification Approaches
- Speech Intelligibility of Dysarthric Speech: Human Scores and Acoustic-Phonetic Features
- Comparison of 5 methods for the evaluation of intelligibility in mild to moderate French dysarthric speech
- Hearing deficits of transformer-based ASR for anechoic and spatial signals
- Disentangled Latent Speech Representation for Automatic Pathological Intelligibility Assessment
- Predicting Intelligibility of Enhanced Speech Using Posteriors Derived from DNN-Based ASR System