This work focuses on daily routine recognition to personalize the hearing aid (HA) configuration for each user. So far, there is only one public data set containing the data of two acceleration sensors taken under unconstrained real-life conditions of one person. Therefore, we create a realistic and extensive data set with seven subjects and a total length of 63449 min. For the recordings, the HA streams the acceleration and audio data to a mobile phone, where the user simultaneously annotates it. This builds the grounds for our comprehensive simulations, where we train a set of classifiers in an offline and online manner to analyze the model generalization abilities across subjects for high-level activities. To achieve this, we build a feature representation, which describes the recurring daily situations and environments well. For the offline classification, the deep neural network, multi-layer perceptron (MLP), and random forest (RF) trained in a person-dependent manner show the significantly best F-measure performance of 86.6%, 87.1%, and 87.3%, respectively. We confirm that for high-level activities the person-dependent model outperforms the independent one. In our online experiments, we personalize a model that was pretrained in a person-independent manner by daily updates. Thereby, multiple incremental learners and an online RF are tested. We demonstrate that MLP and RF improve the F-measure compared to the offline baselines.
This work focuses on sequence learning to improve the daily routine recognition in hearing aids (HA), where the goal is to personalize the device configuration for each user. We apply the sequence methods on two large real-world data sets. One publicly available set contains the acceleration (ACC) data of one person, Huynh, over seven working days, whereas our set includes the real life of seven subjects over 104 days with ACC and audio data of a HA. For both sets, we design statistical features to represent the recurring routine behavior well. In our comprehensive simulations, we analyze several sequence classifiers learning the temporal relationships of high-level activities. The multi-layer perceptron (MLP) and random forest (RF) as an observation model for the hidden Markov model (HMM) show the best F-measure performance of 85.3% and 91.6% on our set and the Huynh set, respectively. In particular, the MLP-HMM combination strongly improves on both sets compared to the non-sequence classifier MLP by 6.7% and 10.2%. Within the segment error analysis, we show that the sequence classifiers improve the temporal prediction stability by a reduction of insertion errors. Thus, the improved sequence classification helps the user to better address his condition due to preferred HA settings.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.