Source-linked AI summary
Trigger Sound Suppression for Misophonia
Vaishnavi Vidyasagar, Jasmine Zhang, Mahima Uliyar, Seunghyun Oh, Emily Catherine Gates, Mark Zachary Rosenthal, Shyamnath Gollakota
TL;DR
Misophonia triggers can cause distress, while common sound-suppression options also silence wanted audio. This paper trains streaming models to remove queried triggers from an acoustic scene and tests the output with listeners with misophonia. Suppressed clips reduced distress and arousal and improved valence relative to raw audio.
Problem
Common coping options either suppress all sound or can cause listening fatigue, while selective suppression of misophonia triggers remains underexplored.
Source: §1 · p. 1Method
The study curates ten trigger classes and trains streaming models with one-hot or multi-hot queries to suppress one to three triggers from the scene.
Source: §1 · p. 1, §3 · pp. 2–3Results
Suppressed clips significantly reduced distress and arousal and improved valence relative to raw audio (p < .001).
Source: §1 · p. 1Takeaways & Limitations
A listening study with 30 adults with clinically elevated misophonia impairment provides human validation of trigger suppression using chewing clips.
Source: §1 · p. 1The dataset uses online audio sources, many recorded close to the microphone, leaving a gap to in-situ reverberant hearable recordings.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Misophonia, a disorder of decreased tolerance to specific sounds, affects 5-20% of the population, yet sufferers have no good options: therapy helps a minority, and earplugs or noise cancellation silence everything. We present a study for neural trigger sound suppression for misophonia, selectively removing trigger sounds. We curate a dataset covering the 10 most common trigger classes. Using streaming dual-path networks operating on 6 ms audio chunks, we explore both one-hot and multi-hot-conditioned models that suppress 1-3 triggers from the acoustic scene. We validate our model outputs in a listening study with 30 adults with clinically elevated misophonia impairment. Participants reported significantly lower distress and arousal, and improved valence, for suppressed audio.
1. INTRODUCTION
Misophonia triggers can cause intense reactions and interfere with daily functioning, while common coping strategies either suppress all sound or have other costs. This paper explores selective trigger suppression and evaluates it with listeners with misophonia.
1. INTRODUCTION
- Earplugs and noise-canceling headphones suppress all sound, while white-noise masking can cause listening fatigue.Selective suppression aims to remove triggers without suppressing the rest of the acoustic scene.Source: §1 · p. 1
- Whether neural target sound extraction can suppress misophonia triggers remains underexplored, with no suitable isolated-trigger training recordings or validation with people with misophonia.Success also needs to be measured by reduced distress, not decibels alone.Source: §1 · p. 1
- In a listening study with 30 adults with clinically elevated misophonia impairment, suppressed clips reduced distress and arousal and improved valence relative to raw audio (p < .001).The study used chewing, a commonly reported primary trigger.Source: §1 · p. 1
- The study curates ten common trigger classes and trains streaming models to suppress one to three triggers using one-hot or multi-hot queries.The models have 0.496M parameters, 10 ms algorithmic latency, and run in real time on an Orange Pi device.Source: §1 · p. 1
2. RELATED WORK
Prior treatments and masking strategies provide limited or non-selective relief, while target sound extraction and low-latency hearables have not been evaluated for misophonia triggers. The paper addresses both trigger separation and distress measurement.
Misophonia and its treatment
- A psychotherapy trial found clinical improvement in 37% of patients, no pharmacological treatment exists, and audiological strategies offer only partial, non-selective relief.Population prevalence estimates range from roughly 5–20%, depending on the instrument and threshold.Source: §2 · p. 1
Target sound extraction
- Prior target sound extraction systems focus on general environmental sounds and have not been evaluated on misophonia triggers.Existing systems use class labels, enrollment audio, or text as target cues.Source: §2 · pp. 1–2
Real-time hearable constraints and clinical validation
- Prior work has not combined evaluation on misophonia triggers with distress ratings, despite hearable latency constraints and the need to measure distress.Delays above 20–30 ms are perceptible.Source: §2 · p. 2
3. METHOD
The method uses queried trigger classes to estimate the residual scene after removing selected triggers. It combines a curated trigger-and-background dataset, synthesized binaural mixtures, and streaming neural separation models.
Problem formulation
- The model outputs the channel-averaged residual signal: the full scene minus the queried triggers.The user specifies the triggers of interest with a multi-hot query.Source: §3 · p. 2
Misophonia trigger dataset
- The curated dataset contains 13,356 clean trigger clips across ten classes, with background clips including speech and music.The 26,742 background clips comprise 5,142 real recordings and 21,600 urban-scene clips; train, validation, and test splits do not overlap.Source: §3 · p. 2
Training mixtures
- Training mixtures combine one to three distinct trigger classes with a background source, using binaural HRIR convolution and trigger-to-background ratios of 0–15 dB.This focuses on triggers at least as loud as the background; quieter triggers are left to future work.Source: §3 · p. 2
Network architecture
- A streaming TF-GridNet takes two-channel binaural mixtures and directly estimates rq[n], using causal convolutions, bidirectional frequency LSTMs, and unidirectional time LSTMs in residual spectral and temporal stages.With LC/LB/LF = 6/6/4 ms at fs = 16 kHz, its algorithmic latency is (LC + LF)/fs = 10 ms.Source: §3 · p. 2
Model variants
- The study compares a single-trigger multi-class model with a multi-trigger model conditioned on multi-hot queries.The multi-hot query is injected into each TF-GridNet block through featurewise linear modulation.Source: §3 · p. 3
4. EXPERIMENTS AND RESULTS
The experiments evaluate trigger suppression against a competing model, examine latency and signal quality, and test affective responses to model output. In a 30-person listening study, suppressed clips were associated with lower distress and arousal and higher valence than original clips.
4. EXPERIMENTS AND RESULTS
- The Orange Pi 5B implementation has 10 ms algorithmic latency; measured processing time is 5.22 ms on average and 5.41 ms at the 95th percentile.The reported processing time depends on the target hardware.Source: §4 · pp. 3–4
- Training uses a composite multi-resolution STFT, L1, and SNR loss; the multi-trigger model is fine-tuned for 50 epochs from the best single-trigger checkpoint.Models are trained in the same chunked mode used at inference.Source: §4 · p. 3
- Sniffling (16.91 dB) and Sneezing (16.47 dB) are separated most reliably, while Drinking (10.73 dB) and Chewing (11.88 dB) are hardest by residual signal SNRi.The passage attributes the difference to broadband, low-energy transients overlapping more with background speech and ambient noise; residual signal SNRi and SI-SNRi differ by under 1.30 dB for every class.Source: §4 · p. 3
- The multi-trigger model outperforms Waveformer across all three trigger-count conditions; Table 3 compares the 0.496M-parameter model with the 2.020M-parameter baseline.Improvements decrease as the number of concurrent triggers increases; the model has the same parameter count as the single-trigger model while handling arbitrary combinations of one to three triggers.Source: §4 · p. 3, Table 3 · p. 3
- Removing lookahead entirely (4/12/0) yields a fully causal 4 ms-latency model but costs 3.70 dB SNRi and 4.21 dB SI-SNRi relative to the 8/4/4 window.Table 4 examines the STFT window budget while keeping its total at 16 ms.Source: §4 · p. 3, Table 4 · p. 3
- For gum-popping suppression, the output tracks the ground-truth temporal envelope and fine structure and preserves the amplitude of non-trigger sounds.Figure 1 shows input mixture, ground truth, and model output waveforms.Source: Figure 1 · p. 4, §4 · p. 4
4.1. Listening Study with Misophonic Individuals
- Distress fell 2.37 points and arousal 2.13 points, while valence rose 1.59 points for model output versus original clips in the 30-participant study.All three paired tests had p < .001; residual distress remained above floor, with participants noting other triggering sounds in the background scenes.Source: §4.1 · p. 4, Table 5 · p. 4
- Participants rated interest at 8.73/10 and perceived impact of losing access at 8.63/10; expected benefit was near 8/10 across four domains.The reported domains were academic/work, home, social, and general.Source: §4.1 · p. 4
5. CONCLUSION
The paper addresses selective suppression of distressing trigger sounds and reports that its latency fits the range allowed by a hearable architecture that reinjects processed sounds of interest. The authors identify dataset realism, trigger coverage, deployment optimization, binaural output, and in-person evaluation as areas for future work.
Discussion
- The model’s latency falls within the range allowed by a hearable architecture that runs ANC on the full acoustic scene and reinjects processed sounds of interest.The cited architecture allows 10 ms algorithmic latency, while ANC itself requires sub-millisecond processing.Source: §5 · p. 4
Limitations and future work
- Online audio sources, many recorded close to the microphone, leave a gap between simulated mixtures and real use.The authors propose in-situ recordings in reverberant rooms using hearable microphones to reduce that gap.Source: §5 · p. 4
- The dataset needs broader trigger coverage.The authors identify expanding coverage to more triggers as future work.Source: §5 · p. 4
- Deployment on hearable-form-factor AI accelerators will require further optimization.The authors point to prior optimization work as a direction.Source: §5 · p. 4
- Future work includes binaural model outputs and end-to-end evaluation with people in person.Source: §5 · p. 4