Fraunhofer Institute for Digital Media Technology

Fragestellung: Wie lässt sich die individuelle Präferenz für Einstellungen der Rauschunterdrückung (noise reduction, NR) bei Hörgerätetragenden schnell und klinisch praktikabel erfassen? Methoden: Mit einer adaptiven kategorialen Skalierung der subjektiv wahrgenommenen [zum vollständigen Text gelangen Sie über die oben angegebene URL]

Electroencephalography (EEG) signals contain rich person-specific information that can be used for biometric applications. This raises both opportunities and concerns in areas such as personalized health care and data privacy. This paper presents an approach for EEG based subject verification, where a transformer-based model learns to generate subject specific embeddings to later determine whether two EEG recordings belong to the same person or not. The proposed method generates subject-specific embeddings using pre-trained models and performs subject verification using a distance-based matching strategy with threshold tuning. To assess the models generalization, cross-session and cross-dataset experiments are conducted using two EEG datasets with different recording montages with one being made up of subjects with medical pathologies. Results indicate high performance even when trained and evaluated across different subject groups, and recording configurations, reaching accuracies above 80%. These findings show that identity-related information can be extracted from EEG signals independent of health status or recording protocols, while also highlighting the need for privacy preservation of EEG data to protect sensitive subject-specific information.

2148

Decline in vigilance due to fatigue is a common concern in traffic safety. Partially automated driving (PAD) systems can aid driving but decrease the driver's vigilance over time, due to reduced task engagement. Mobile EEG solutions can obtain neural information while operating a vehicle. The purpose of this study was to investigate how the behavior and brain activity associated with vigilance (i.e., alpha, beta and theta power) differs between PAD and manual driving, as well as changes over time, and how these effects can be detected using two different EEG systems. Twenty-eight participants performed two 1-h simulated driving tasks, while wearing both a standard 24 channel EEG cap and a newly developed, unobtrusive and easy to apply 10 channel mobile EEG sensor-grid system. One scenario required manual control of the vehicle (manual) while the other required only monitoring the vehicle (PAD). Additionally, lane deviation, percentage eye-closure (PERCLOS) and subjective ratings of workload, fatigue and stress were obtained. Alpha, beta and theta power of the EEG as well as PERCLOS were higher in the PAD condition and increased over time in both conditions. The same spectral EEG effects were evident in both EEG systems. Lane deviation as an index of driving performance in the manual driving condition increased over time. These effects indicate significant increases in fatigue and vigilance decrement over time while driving, and overall higher levels of fatigue and vigilance decrement associated with PAD. The EEG measures revealed significant effects earlier than the behavioral measures, demonstrating that EEG might allow faster detection of decreased vigilance than behavioral driving measures. This new, mobile EEG-grid system could be used to evaluate and improve driver monitoring systems in the field or even be used in the future as additional sensor to inform drivers of critical changes in their level of vigilance. In addition to driving, further areas of application for this EEG-sensor grid are safety critical work environments where vigilance monitoring is pivotal.

Objective This study investigated a previously evaluated self-adjustment procedure with respect to its applicability for asymmetrical hearing loss (AHL). Self-adjusted settings were evaluated for speech recognition in noise and sound preference. Design Participants were given the possibility to adjust the left and right hearing aid separately using a two-dimensional user interface. Two different adjustment sequences were tested. Realistic everyday sound scenes in a laboratory environment were presented. The difference between the ears regarding their speech recognition in noise was tested with two spatial conditions, unaided as well as with the prescriptive formula and the self-adjusted setting. Study sample Nineteen experienced hearing aid users (median age 76 years) with different degrees of AHL were invited to participate in this study. Results Participants adjusted a higher gain slope across frequency in the worse ear than in the better one. The two adjustment sequences resulted in significantly different adjustment durations and gain settings. The difference between the ears regarding speech recognition in noise did not change with the self-adjustment. Overall, group-mean effect sizes were small compared to the parameter space. Conclusions The adjustment procedure can be used also by hearing aid users with AHL to find a possibly preferred gain setting.

Deep audio representations, also known as embeddings, recently became a popular alternative to conventional features like spectrograms for a wide range of audio classification tasks because of their domain-agnostic character and reduced training costs. Still, the usage is often limited to rather computationally intensive system due to the nature of their extraction from large networks. This paper aims to minimize the computational costs of embedding extraction by distilling the knowledge of the OpenL3 audio network to a smaller student network. Results show that the student network maintains comparable performance as the teacher network on various music and ambient noise classification tasks, while reducing the network size by over 90\% and the computational load by five times.

In this paper, we introduce FLCrypt <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> For evaluation access to FLCrypt, please contact the main author or reach us via https://www.idmt.fraunhofer.de/en/contact.html., a library designed to enhance Federated Learning with additional privacy guarantees by applying Fully Homomorphic Encryption to the model aggregation stage, thereby preventing the aggregator from accessing the unencrypted model parameters of the training participants. We evaluate our approach by comparing its accuracy against an unencrypted baseline using an audio classification task aimed at distinguishing metal balls with different surfaces, which serves as a proxy for fault detection in industrial sound analysis. Our findings indicate a marginal decrease in accuracy due to applying Fully Homomorphic Encryption, alongside a significant increase in both runtime and memory demands. Our analysis also concludes that runtime increases linearly with the number of model parameters. Our results lead us to affirm the viability of FLCrypt for Federated Learning applications in acoustic sensor networks with elevated security requirements such as sound classification.

This paper introduces a novel multi-task transformer for detecting synthetic speech. The network encodes magnitude and phase of the input speech with a feature bottleneck, used to autoencode the input magnitude, to predict the trajectory of the first phonetic formants (F0, F1, F2), and to distinguish whether the input speech is synthetic or natural. The approach achieves state-of-the-art performance on the ASVspoof 2019 LA dataset with an AUC score of 0.932, while ensuring interpretability at the same time.