We introduce a new, open-source framework for training a low-resource, low-latency ML hearing aid, based on a loss function that compares neural activity patterns from a differentiable model of the human inner ear.
EdgeConvGRU is a lightweight, 34K-parameter neural network architecture that enables accurate, real-time acoustic scene classification tailored for resource-constrained hearing aid devices.
This work presents a prototype of a web-based AI voice agent for speech communication testing and shows that, while natural low-latency interaction is feasible, future studies will require more sensitive task metrics.
Accounting for pre-cochlear implant (CI) cognitive and linguistic abilities improves prediction of post-CI speech recognition outcomes.
Routine occupational audiometry combined with machine-learning models can stratify airport catering workers by high-frequency hearing risk and support targeted hearing-conservation strategies.
A neural network trained on physiologically simulated data can decode cochlear damage profiles from non-invasive recordings, even in ears with identical audiograms.
A machine learning model optimized for everyday hearing tasks exhibits human-like psychoacoustic thresholds, suggesting perceptual sensitivity is limited by efficient representations of sound rather than internal noise.
DeepFS4 introduces a novel end-to-end deep learning architecture that integrates speech enhancement directly into the FS4 sound coding strategy by applying noise reduction on the temporal envelope while maintaining the temporal fine structure for the most apical channels, significantly improving speech intelligibility in noise for cochlear implant users without adding latency.
This study applies traditional statistical analysis and machine learning to evaluate Frequency Following Response (FFR) patterns in children with school difficulties or stroke.
We present a UMAP-based Hearing Loss Map that identifies novel sensorineural hearing loss phenotypes and longitudinal trajectories from a large clinical dataset, going beyond conventional audiogram-based classification.
Computational closed-loop hearing compensation can personalize signal processing for both standard and hidden hearing loss, improving objective intelligibility outcomes while preserving perceived listening quality.