Expected Creak and Misplaced Creak: Separating Linguistic Phonation from Depression in Mandarin

Mandarin speakers routinely produce creaky voice, a low, irregular buzz of the vocal folds that most people know as "vocal fry." In Mandarin this is mostly not a sign of anything wrong. Creak is part of how the low Tone 3 is pronounced, and it reliably marks the end of a phrase. Depression research, meanwhile, treats vocal-fold irregularity as a symptom: in English clinical speech, shimmer and breathiness rise with depression severity and with psychomotor retardation, the clinical term for slowed movement and speech. Speech-based depression detectors for Mandarin pool all of this into a single acoustic signal. They cannot tell whether a patient creaks more because the larynx is less well controlled, or because the patient pauses more often and so produces more phrase endings where creak is normal anyway.

We propose to model where creak should occur in a fixed read passage, given each syllable's tone and its position in the phrase, using healthy speakers. Each speaker's creak is then split into an expected part and a misplaced part. The same logic applies to tone sandhi, the rule that turns the first of two adjacent Tone 3 syllables into a rising tone. We measure how much each speaker's sandhi rise falls short of that speaker's own ordinary rising tone. These context-controlled measures are then correlated with standardized depression scores (PHQ-9) in the read-speech portion of the MODMA clinical corpus. The read passage is identical for every participant, so this design avoids the interviewer-prompt shortcut that inflates results in interview corpora. The scientific question is a dissociation: is depression-related creak in Mandarin a phrasing and breathing effect, a laryngeal-control effect, or both?

Core Claim

The claim is that a speaker-level "misplaced creak" rate predicts PHQ-9 severity after controlling for pause rate. Misplaced creak means creak on non-low tones in phrase-medial positions, where healthy Mandarin speakers rarely creak. A context-blind global creak rate either fails this test or passes it only because it carries boundary frequency with it. A second, weaker claim concerns sandhi. Severity tracks the size of a speaker's sandhi Tone 3 rise relative to the same speaker's lexical Tone 2 rise. Taking this ratio cancels the overall pitch-range flattening that depression is already known to cause, so any remaining correlation points to a deficit in executing the sandhi rule itself.

Either outcome of the dissociation is informative. Suppose global creak correlates with severity but misplaced creak does not. Then depression-related creak in Mandarin is largely a by-product of more frequent phrase breaks and longer breath groups. That reading fits the finding of Creak in the Respiratory Cycle that creak appears near the end of an exhalation, when lung volume runs low, as an air-saving strategy. Suppose instead that misplaced creak survives the pause-rate control. Then there is a laryngeal signal that the global measures were blurring together with the phrasing effect.

What Prior Work Establishes and Leaves Open

The expectation model rests on two healthy-speech findings. Creaky Voice as a Function of Tonal Categories and Prosodic Boundaries analyzed six speakers of the COSPRO_01 corpus. It showed that creak is concentrated in Tone 3 and the neutral tone, rarest in Tone 1, and that its probability rises strictly with boundary strength, ranked from weakest to strongest as B2 < B3 < B4 < B5. At B5, the utterance-final boundary, syllables creak regardless of tone. That paper gives us the context structure but no clinical variables, and it could not model sandhi syllables separately because there were too few of them. Mandarin Tone Sandhi Realization: Evidence from Large Speech Corpora used about 2,300 speakers to show that sandhi Tone 3 has a smaller pitch rise than lexical Tone 2, so the rule is gradient rather than all-or-nothing. It also showed that the rise varies with sex, age and region. That paper measured pitch only, and its demographic effects tell us which nuisance variables we must partial out.

Two papers shape how we measure creak. Acoustic Differences Between English /t/ Glottalization and Phrasal Creak showed that boundary creak and abrupt glottal closures differ in kind. Boundary creak has static low pitch and a low $H_1^*-H_2^*$, a spectral-tilt measure of vocal-fold tension. Abrupt closures show a dynamic drop in cepstral peak prominence (CPP), a measure of how periodic the voice is. We therefore record both static and dynamic creak cues rather than a single creak flag. The role of creaky voice in Mandarin tone 2 and tone 3 perception found that creak at the pitch dip shifts listeners toward hearing Tone 2 and delays the perceived turning point. Creak and turning-point timing therefore interact, and sandhi pitch measures must be robust to creak. Revisiting Pitch Jumps: F0 Ratio in Seoul Korean adds that sudden pitch jumps during creak can be real rather than tracking errors. For that reason we will hand-check the pitch tracks rather than smoothing the jumps away.

On the clinical side, Vocal-source biomarkers for depression: a link to psychomotor activity supplies the clinical motivation. In English patients it found that shimmer and lower harmonics-to-noise ratio (HNR) track HAM-D, QIDS and the psychomotor-retardation subscales, while jitter correlated only weakly (r ≈ 0.11). It had no tonal or prosodic context. Speech Features for Depression Detection observed that depression can sound creaky or breathy depending on the individual. That is a warning that a single direction of effect may not hold across speakers.

On Mandarin speech, ECAPA-TDNN Based Depression Detection from Clinical Speech and Depression Detection in Speech Using Transformer and Parallel Convolutional Neural Networks classify depression from MFCCs or spectrograms with high accuracy, but they offer no account of which speech behavior carries the signal. DAIC-WOZ: On the Validity of Using the Therapist's prompts in Automatic Depression Detection from Clinical Interviews constrains our design. It showed that interview-based models can reach F1 of 0.90 by exploiting the interviewer's questions, which is why we anchor on read speech.

Finally, Evaluating the Examiner: The Perils of Pearson Correlation for Validating Text Similarity Metrics shows that Pearson r is distorted when scores cluster. That applies directly here, because PHQ-9 scores in a case-control sample split into a low cluster for controls and a high cluster for patients.

Research Program

Data. The primary data are the "North Wind and the Sun" reading recordings from MODMA dataset: a Multi-modal Open Dataset for Mental-disorder Analysis. The corpus has 52 Mandarin speakers: 23 diagnosed with major depressive disorder (MDD) using MINI and DSM-IV, and 29 controls, all with PHQ-9 scores. Its picture-description task serves as a secondary condition with no interviewer prompts but variable text.

Annotation. The text is fixed, so we transcribe its tones and sandhi sites once. We will count the Tone 3 + Tone 3 sites in the passage before committing to the sandhi analysis; if they are few, sandhi becomes secondary and the picture descriptions supply additional tokens. Forced alignment gives syllable boundaries. MODMA has no break-index labels, so realized boundaries will be labeled with a pause-and-lengthening detector. The FPCA contour classifier of Tonal coarticulation as a cue for upcoming prosodic boundary, which reported 92% accuracy on COSPRO, will be used as a cross-check; we expect it to lose accuracy when moved from Taiwan to Gansu Mandarin. Creak will be detected frame by frame with a residual-plus-pulse-regularity detector of the kind used by Kuang. We will add VoiceSauce-style $f_0$, CPP, $H_1^*-H_2^*$ and $H_1^*-A_n^*$, and validate the detector against hand labels on a stratified 10% subset.

Expectation model. For each syllable $s$ we fit the probability of creak $p(s)$ with a logistic model. Its inputs are tone category, sandhi status, boundary strength, distance to the end of the breath group, and speaker sex. We fit it on healthy controls only, leaving out the speaker being scored each time so that no speaker contributes to their own expectations. Each speaker's misplaced-creak score is

$$M = \frac{1}{|L|}\sum_{s \in L} \big(c_s - p(s)\big),$$

where $c_s$ is 1 if syllable $s$ creaks and 0 otherwise, and $L$ is the set of syllables in low-expectation contexts: non-Tone-3, non-neutral-tone, phrase-medial syllables. A parallel score $E$ over high-expectation contexts measures whether expected creak is under- or over-produced. Pause rate, or the number of boundaries per second, is recorded as a separate variable. The sandhi undershoot score is $U = R_{\text{sandhi}} / R_{T2}$, where each $R$ is the speaker's median log pitch-rise range on that syllable type, following the rise measure of Wang et al.

Statistics. The primary analysis is a pre-registered Pearson correlation between $M$ and PHQ-9, with $U$ as the secondary measure. Before computing r we transform each acoustic score toward normality and check it. We then compute partial correlations controlling for sex, age, speech rate and pause rate, with bootstrap confidence intervals. Spearman's rank correlation is reported alongside. Because the case-control split clusters PHQ-9, we also report the correlation within the MDD group alone; this is the analysis least vulnerable to the distortion that the Pearson-critique paper describes.

The power budget is honest. With n = 52, Fisher's z transform puts the detection threshold for 80% power at a two-sided α of 0.05 near r ≈ 0.38, from $\tanh(2.80/\sqrt{49})$. Within the 23 patients only effects near r ≈ 0.55 are reliably detectable. We therefore treat the within-group analysis as supporting evidence, not as a stand-alone test.

Baselines

The naive reference predicts PHQ-9, or the diagnostic group, from sex, age and speech rate alone. Speech rate matters because slowed speech is the most obvious motor sign, and any phonation measure must add something beyond it. All feature-based models are fit under the same leave-one-speaker-out protocol on the same reading recordings. Beating the naive reference means a significant gain in explained variance, tested by nested-model likelihood ratio, together with a higher leave-one-out area under the ROC curve.

The classical voice-quality baseline computes jitter, shimmer, HNR, mean CPP, $f_0$ variance and a context-blind global creak rate over the same reading passages. This follows the feature set of the vocal-source biomarker study, which is the strongest interpretable competitor because it already links phonation to depression severity. No Mandarin number exists for it; the English reference point it supplies is jitter versus HAM-D Total at r ≈ 0.11 in Vocal-source biomarkers for depression: a link to psychomotor activity. Beating it means that $M$ correlates with PHQ-9 above the baseline's best single feature, and that $M$ retains a significant partial correlation once the baseline features and pause rate are in the model.

The strongest existing method is end-to-end spectrogram classification. On MODMA audio, Depression Detection in Speech Using Transformer and Parallel Convolutional Neural Networks reports 95.8% unweighted accuracy and 96.7% F1. The file does not say whether that result is speaker-independent or restricted to reading speech, so we will not compare against it as if it were. We will instead retrain an ECAPA-TDNN in the style of ECAPA-TDNN Based Depression Detection from Clinical Speech, which reported 90.4% speaker-level accuracy on its own hospital corpus. The retrained model uses only the MODMA reading audio, with leave-one-speaker-out folds and the same hyperparameter search budget as our feature models. We do not expect five interpretable scores to beat it on accuracy. The bar is that our scores come within a few points of the network's area under the curve under identical folds, and that they explain which behavior carries the signal. As a probe of the network, we will test whether its outputs correlate with $M$ or with pause rate.

The key ablation swaps $M$ for the context-blind creak rate. A second ablation fits the expectation model without boundary terms, which isolates whether conditioning on boundaries is what matters.

Why not just…?

Why not just use a global creak rate, since everyone reads the same text? This is the most serious threat. A fixed text already balances tone composition across speakers, so tone context is controlled before any modeling. What a fixed text does not control is phrasing. Speakers decide where to pause, and patients who pause more create more phrase endings where creak is expected. The global-rate ablation, with and without pause rate partialed out, settles the question directly. If global creak does as well as $M$ and loses its effect once pause rate is controlled, we will report that depression creak in this sample is a phrasing effect. That would still be a substantive result, and it would mean the misplaced-creak score adds no sensitivity.

Why not just train a deep model, given that accuracy on MODMA is already above 90%? With 52 speakers, any accuracy figure has wide confidence intervals. Spectrogram models also give no account of whether the signal is breathing, phrasing or laryngeal control, which is exactly the question here. The retrained ECAPA baseline under identical folds shows how much accuracy the interpretable route gives up. The probe correlation shows whether the network is quietly tracking pause rate.

Why not just use Spearman correlation or a nonlinear regression instead of Pearson? Spearman is reported for every test. Pearson remains primary because the partial-correlation and nested-regression machinery we need is linear, and we transform and check the scores before using it. The real risk is the clustering of PHQ-9 across the two groups, not nonlinearity as such. The within-MDD analysis and the bootstrap intervals address that risk directly.

Decisive Checks

The first check is whether $M$ correlates positively with PHQ-9 at the pre-registered threshold and keeps that correlation after partialing out pause rate, speech rate and sex. Evidence for a laryngeal signal requires that the context-blind creak rate, under the same controls, does not. The second check is whether the sign of the effect replicates in the picture-description task, which has different text and no interviewer prompts. The third check is whether $U$ declines with severity while the raw Tone 2 rise does not, which would separate a sandhi-execution deficit from general pitch flattening. If none of these hold, the conclusion is that functional creak and sandhi in Mandarin read speech are spared by depression. That is itself a useful negative result for anyone building phonation-based screens for tonal languages.