Properties (17)
Id 773849c0-7964-47f3-b476-9dd2dcbc473c
Year 2025
Abstract
Cssclasses
  • purpleRed
Aliases
  • (Clonan et al., 2025)
Import Date 06-07-2026
Template Version Notes •corrected error in rendering the linked PDF in metadata section • added the tag “noted” to the definition of comments to ignore ( like with the “in another note” tag. this is so that when making notes on chapters of a textbook, i can still export things for manual addition to the obsidian note without having to keep track of what has already been added ). Also moved the version notes to the yaml section • this version added the bookmarks section to auto collect all bookmarks. it also dropped the template auto fill of the initial source field because we never got it working and just overwrote it manually anyway. this version also added logic to handle “in another note” tag. annotations marked with this are completely ignored. used when i have information covered in a bigger note, but still want to highlight that part in text for visual clarity (so as to avoid duplicates of the same information) • running the import more then once will overwrite the file, so do not import again if you have made changes in obsidian that are not in the zotero article.
Dv Authors Alex C. Clonan, Ian H. Stevenson, Monty A. Escabí,
Dv Initial Source Misophonia MOC
Dv Status #status/reviewed
Dv Limitations • Small participant sample and limited training sounds in FOAMS. • FOAMS samples only a small subspace of possible natural sounds. • Accuracy of pTFs and out-of-set predictions requires verification. • Midbrain model use does not establish a causal subcortical link. • Low-dimensional visualizations may not fully capture high-dimensional categorization.
Dv Future Directions • Perceptual studies using synthetic sounds and more acoustically diverse sound sets. • Larger populations to identify population-level transfer function differences and set diagnostic standards. • Design optimized misophonia corpora balanced with misophonic and non-misophonic sounds. • Extend GPR with electrophysiological, behavioral, and multisensory predictors (hybrid multimodal model). • Combine GPR with symptom-dimension mapping via MIMO regression. • Real-time interventions that identify and attenuate triggers.
Dv Relevance • a very interesting paper for methods around modeling the acoustics of Misophonia and using these models to detect triggers in pre-existing audio. • study used a statistical method of modeling acoustic features to show the feasibility of using such models to predict misophonia aversiveness to sounds, both tested and untested. • paper has an interesting interaction with a previous non Misophonia Kumar paper and the andermanePhenomenologicalCartographyMisophonia2023\(Andermane et al., 2023) paper • study introduces a the concept of individualized transfer functions, a measure of the spectrotemporal modulations that contribute to individual sound aversion. the authors claim this can be computed to give perceptual ratings fit from or predicted for any sound once calculated.
Checked For Zettle Notes true
Zotero Note Link Generated true
Notes Audited 2026-06-07T16:58:00
Audit Notes audited as review was done

•

Identifying the acoustic fingerprints of trigger sounds and predicting discomfort for misophonia

Status:: #status/reviewed

Metadata

Abstract
Background/Purpose Background-
• Misophonia evokes severe anxiety, anger, or disgust to commonly encountered everyday sounds.
• Classified as a disorder of reduced tolerance and altered emotional perception of sounds.
• Conflicting evidence on auditory cortex involvement; altered functional communication observed between auditory cortex and saliency/emotion regions (insula, amygdala, prefrontal cortex, hippocampus).
• Auditory midbrain sensitivity has been correlated with tolerance to background noise.
• Trigger sounds vary widely person-to-person; no consensus on precise acoustic features driving the disorder.
• Existing diagnostics rely on categorical sound preferences and listening profiles.
• No current diagnostic tool pinpoints specific acoustic properties triggering events or predicts aversion to novel sounds.
• Substantial heterogeneity exists in specific triggering sounds across people, even within broadly aversive categories.


Purpose-
• Propose a bottom-up auditory modeling approach to characterize and diagnose misophonia preferences and predict tolerance ratings.
• Determine whether bottom-up statistical features from periphery and midbrain explain sound aversion.
• Identify specific spectrotemporal features underlying aversion for individual listeners.
• Demonstrate extrapolation to untested sounds and personalized trigger detection in continuous audio.
Participants/samples • FOAMS corpus; 21 listeners total.
• Modeled n = 19; 2 excluded for reporting no aversion.
• n = 4 classified misophonic; n = 15 non-misophonic.
• Participants 1, 3, 5, 6 exceeded S-Five threshold; participants 5, 6 exceeded DMQ impairment threshold.
Measures • Duke Misophonia Questionnaire (DMQ); summed scores include symptoms, coping, impairment, belief factors.
• S-Five questionnaire; summed scores include internalizing appraisals, externalizing appraisals, sense of emotional threat, outbursts, impact factors.
• Tolerance/discomfort ratings (0–5; 0 = benign, 5 = aversive).
Procedures • Hierarchical Framework: The study utilizes a biologically inspired model mimicking the cochlea and auditory midbrain, decomposing sounds into spectrotemporal modulation components to create a midbrain-modulation power spectrum (mMPS). (page 2)

• FOAMS: 82 potential trigger sounds; participants rate discomfort 0–5; 4 sounds per category.
• Cochlear stage: gamma-tone filterbank with equivalent rectangular bandwidths (82 filters, 20–20k Hz, 50% overlap), hair-cell rectification, 750 Hz lowpass (auditory nerve), nonlinear output compression; produces cochleogram.
• Midbrain stage: bank of log-spaced spectrotemporal modulation filters (1-octave bandwidths); temporal modulation ±0.5 to ±512 Hz, spectral modulation 0.25 to 8 cycles/oct; produces midbrainogram.
• mMPS: time-averaged power across midbrain outputs; 82 frequencies × 38 temporal modulations × 8 spectral modulations.
• GPR: PCA dimensionality reduction (20 components for FOAMS, 40 for ESC50), GLM fit by maximizing log-likelihood, 8-fold cross validation; weights projected back to mMPS space to form pTFs.
• Diagnostic relation: GPR intercept and gain compared against summed DMQ and S-Five scores.
• Convolutional trigger detection: 2-second window, 1-second overlap; time-varying mMPS, PCA, GLM weights applied.
• Detection tested on a 10-minute segment from the movie 'Up' with 18 inserted FOAMS triggers (9 generally aversive, 9 conditionally aversive) at 0 dB SNR.
• Hierarchical clustering of raw tolerance scores (Ward's minimum variance, MATLAB 'linkage').
• ESC50 extrapolation; t-SNE on mMPS features; YAMNet for category classification.
Results
Model prediction
• mMPS summary statistics predict individual tolerance ratings across acoustic categories.
• Participants show markedly different average tolerance across categories.
• GPR captures distinct aversion profiles (general low aversion; sharp aversion to bodily/respiratory sounds; aversion to flipping paper and typing).
• Model predicts perceptual variance well above a random worst-case baseline.

Transfer functions
• Each participant has a unique pTF morphology.
• A common low-frequency, slow temporal-modulation component drives aversive tolerance across most listeners.
• Low-aversion participant has a weaker (low-gain) transfer function.
• Faster temporal modulations and high frequencies drive aversion in the paper/typing-sensitive participant.

Diagnostic correlation
• GPR intercept positively associated with summed DMQ and S-Five scores.
• GPR gain positively associated with summed DMQ and S-Five scores.
• Intercept and gain associated with upper-quartile FOAMS aversion scores.
• Intercept and gain associated with each other.

Clustering
• Hierarchical clustering of raw scores reveals distinct participant clusters.
• Clusters differ in sensitivity (general low aversion; high aversion across most categories; aversion specific to breathing and oral sounds).
• Cluster-average transfer functions show distinct spectrotemporal patterns matching individual pTFs.

Extrapolation
• GPR predicts tolerance for out-of-sample ESC50 sounds.
• Predictions for shared categories (breathing) consistent across datasets.
• Novel categories (snoring, crow) predicted aversive for some clusters.
• Within-category heterogeneity in predicted aversion (some breathing exemplars not aversive; some typing exemplars aversive).
• Out-of-distribution predictions maintain personalized preferences.

Acoustic space
• t-SNE shows some categories cluster acoustically (machine sounds, vocalizations) while others spread (bodily sounds).
• FOAMS sparsely samples the acoustic space relative to ESC50.
• Within-category diversity more evident in ESC50 than FOAMS.
• Projected predictions show participant differences in overall severity and aversive regions.

Trigger detection
• Convolutional detection identifies injected generally-aversive triggers across participants and time.
• Conditionally-aversive sounds show participant-specific rather than reliable activation.
• Some unplaced audio segments are predicted aversive.
• Mixture predictions correlate with clean aversion responses.
• YAMNet category timing largely uncorrelated with predicted tolerance; high-discomfort regions mostly classified "ambiguous".
Discussion/Limits Conclusion-
• Time-averaged summary statistics from a biological auditory model suffice to predict sound tolerance trends across participants with distinct aversion patterns.
• Acoustic features, spectrotemporal modulations in particular, can characterize individualized patterns of aversion without explicit categorization.
• Individual modulation cues have small impact, but co-occurring cues build into ensembles with high discomfort.
• A feature-based model agnostic to semantics can identify misophonic etiologies and has diagnostic potential.
• Use of a midbrain model does not causally link findings to subcortical or midbrain computations.
• Unlike prior group-average approaches, GPR generates individualized models of each listener's profile.
• pTF parameters correlate with established diagnostic scores, suggesting diagnostic use.
• The pTF can be conceptualized as a multi-dimensional psychometric function for natural sounds.
• Separating sound representation from behavioral prediction makes GPR highly interpretable.
• The sound-computable approach generalizes to novel triggers and operates in continuous sound mixtures.
• pTFs account only for acoustic contributions; visual, emotional, and attentional factors also contribute.
• Authors raise, as an open question, whether misophonic perception is driven by conditioned ensembles of acoustic features associated with emotional context.
Limitations:: • Small participant sample and limited training sounds in FOAMS.
• FOAMS samples only a small subspace of possible natural sounds.
• Accuracy of pTFs and out-of-set predictions requires verification.
• Midbrain model use does not establish a causal subcortical link.
• Low-dimensional visualizations may not fully capture high-dimensional categorization.
Future Directions:: • Perceptual studies using synthetic sounds and more acoustically diverse sound sets.
• Larger populations to identify population-level transfer function differences and set diagnostic standards.
• Design optimized misophonia corpora balanced with misophonic and non-misophonic sounds.
• Extend GPR with electrophysiological, behavioral, and multisensory predictors (hybrid multimodal model).
• Combine GPR with symptom-dimension mapping via MIMO regression.
• Real-time interventions that identify and attenuate triggers.
Relevance:: • a very interesting paper for methods around modeling the acoustics of Misophonia and using these models to detect triggers in pre-existing audio.
• study used a statistical method of modeling acoustic features to show the feasibility of using such models to predict misophonia aversiveness to sounds, both tested and untested.
• paper has an interesting interaction with a previous non Misophonia Kumar paper and the (Andermane et al., 2023) paper
• study introduces a the concept of individualized transfer functions, a measure of the spectrotemporal modulations that contribute to individual sound aversion. the authors claim this can be computed to give perceptual ratings fit from or predicted for any sound once calculated.

Notes

key terms and concepts

my engagement notes

my questions

my comments

Bookmarks

reading notes

image-5-x36-y95.png Fig. 1. Representing natural sounds using auditory model summary statistics. A) Among all possible sounds, only a subset act as misophonia triggers with identifiable category labels. Here we aim to develop a framework for identifying which sounds act as triggers (region outlined in orange) for individual people. B) We fit behavioral data from the FOAMS dataset, where participants both with and without misophonia rated discomfort (0–5) for a set of selected sound stimuli. C) To characterize the differences among sounds, we use a neurally-inspired, midbrain modulation power spectrum. Sounds are first decomposed using a bank of cochlear filters. The color scale is proportional to the sound power at each time and frequency bin (blue=low power; red and yellow=high power). The cochleogram is then decomposed using a second stage of modulation filters. Averaging over time, we have a representation of the sound statistics based on the power of specific temporal and spectral modulations in specific frequency-channels. D) Stimuli differ substantially in mMPS, and these differences can be used to predict differences in discomfort ratings. page 5

image-6-x21-y50.png Fig. 2. Fitting sound-computable models of perceptual ratings with generalized perceptual regression (GPR). Sounds are first represented using the mMPS (left). Since the midbrain representation is quite high-dimensional, we use PCA to reduce the sounds to a low-dimensional set of components. We then fit a generalized linear model (GLM) that predicts aversion scores using the low-dimensional acoustic embeddings as predictors. To interpret the model fit, we project the coefficients of the GLM back into the high-dimensional space to give a “transfer function”. Positive values in the transfer function (red) indicate that power in this specific frequency and modulation band is associated with higher aversion scores, while negative values (blue) indicate that power in this band is associated with lower aversion scores. page 6

image-7-x19-y165.png Fig. 3. Predicting individual participant sound aversion ratings through generalized perceptual regression (GPR). A) Here we show the model’s k-fold cross-validated predictions as a category averaged aversion score (average of 4 sounds per category and trials per sound) for selected participants (Participant = 13, 1, ,3, 19, 6). Participants were organized according to hierarchical clustering presented in Fig. 5. Red bars show the model prediction; white dots denote participant responses. Red participant labels indicate they met diagnostic criteria for misophonia with either the DMQ and/or S-Five questionnaire. B) Here we show the model’s k-fold cross-validated predictions comparing the predicted value with the actual reported aversion as a scatterplot. Notably, the model can account for different aversive profiles (i.e., Participant 13 is not sensitive to many sounds, while Participant 6 is sharply aversive to bodily/respiratory sounds). C) Distribution of variance explained (r2) values across the whole population (n = 19). D) Scatter plot showing the cross-validated actual versus predicted tolerance ratings across the whole population (n = 19). E) The perceptual transfer functions of each participant indicate the sensitivity in different frequency and modulation bands that increase (red) or reduce (blue) aversion scores. Notably, each participant has a unique transfer function morphology, suggesting that a unique set of spectrotemporal features influence each listener’s tolerance to sounds. Transfer functions are shown for the 0 cycles/octave spectral modulation which dominate each participants sensitivity. F) The participant average transfer function exhibits a low frequency (<1.6 kHz) and slow (<2 Hz) component that is common to most listeners, and which drives aversive tolerance (red hot spot). page 7

image-9-x36-y109.png Fig. 5. Hierarchically clustering participant data. To determine sensitivity scores for each individual sound in A) we averaged across given sound exemplars across two trials and hierarchically clustered the raw scores (left matrix) by sound exemplar (x-axis) and participant number (y-axis). Similarity can be seen through the dendrograms on each axis. The model prediction is mapped into the same space at 20 principal components (right). Red participant labels indicate they met the criteria for either the DMQ and/or S-Five questionnaire. B) To assess whether these clusters presented themselves in our acoustic subspace, we then plotted the average transfer function for each cluster (C1-C5). Notably, the clusters differ between one another yet maintain similarity with their exemplars shown in Fig. 3 C) To test how well our model can extrapolate to novel data, we used a novel test set from the ESC50 corpus. To do so, we generated mMPS vectors for ESC50 corpus sounds and performed PCA with the covariance matrix generated in Fig. 2. We then generated a novel set of 2000 predictions and averaged across 50-categorical conditions (which have 40 trials each) (top). D) We then show the individual trial predictions within two selected sound categories that are both present in ESC50 and the FOAMS corpus, Breathing (bottom left) and Keyboard Typing (bottom right). While most breathing sounds elicit a higher predicted aversion score, there are a few exemplars which present low aversion scores. Similarly, in the keyboard typing condition, most sounds elicit a low aversion score while a few exemplars appear to be highly aversive based on their acoustics. page 9

image-15-x116-y146.png Fig. 8. Acoustic subspaces may allow for more targeted and specific identification of personalized misophonic triggers. A) Psychoacoustically, participants can be sensitive to different sets of categorical trigger sounds. Individual participants can have divergent acoustic subspaces (participant one in red, sensitive to breathing/ typing; participant two in purple, sensitive to chewing/typing). Since the GPR model can predict different participant specific sensory aversions from the acoustic space, its plausible that individual participants are sensitive to different ensembles of acoustic features. B) Current sound corpora treat misophonia as a categorically determined or top-down influenced. Characterizing misophonic trigger sounds by their acoustics may explain within-category variability (Ex1/Ex2, green/blue) not explainable by semantic labels. These subspaces may identify possible triggers in classically non-misophonic sounds (red, crow and snoring), which have comorbid acoustic features (orange) with implicated misophonic triggers (green, blue). Similarly, these possible identified trigger sounds are likely exemplar-specific. page 15

Misc notes

contextual statement

In a 2025 study, Clonan attempted to predict misophonia discomfort ratings and identify the acoustic features driving sound aversion by deriving auditory midbrain summary statistics (mMPS) from FOAMS sounds and fitting individualized generalized perceptual regression models for 19 listeners assessed with the DMQ and S-Five.

Generalized Perceptual Regression (GPR)

(pdf)

perceptual transfer functions

contrasting with a previous (non miso) Kumar 2008 paper on sound decomposition (pdf) , and the use of method in Andermane 2023 showing null effects in a Misophonia study

the authors note and address two possible issues to the novelty of their findings

the authors address this with the following arguments

Kumar's use of partial least squares (PLS) vs the current methods use of dimensionality reduction (pdf)
previous studies used average responses, this study used individualized responses (pdf)
this studies unique extensions of previous finings (pdf)

Zettelkasten

Generalized perceptual Regression framework applied to Misophonia

Clonan 2025 contrasting with a previous (non miso) Kumar 2008 paper on sound decomposition , and the use of method in Andermane 2023 showing null effects in a Misophonia study

predicting sound tolerance via acoustic summary statistics in Misophonia

Enter your search text in the box above

Select a result to preview