From: 2025-06-02 14:00 to 15:00
Place: Large Conference Room 2485-88 in the M-building, LTH
Contact: bo [dot] bernhardsson [at] control [dot] lth [dot] se

Date & Time: June 2nd, 14:00-15:00
Location: Seminar Room M 3170-73 at Dept. of Automatic Control, LTH
Author: Cajsa Thulin & Josefine Müller
Title: The Effect of SNR on Auditory Attention Decoding
Supervisor: Bo Bernhardsson
Examiner:  Pontus Giselsson

Abstract: This study investigated how acoustical challenges, specifically changes in signal-to-noise ratio (SNR), affects the neural tracking of attended and ignored speech in individuals with hearing impairment. The analysis began with the speech envelope and was progressively extended by incorporating phonetic information as well as outputs from all layers of OpenAI’s Whisper model. The goal being better understanding of the brains’ audio processing pathways during Cocktail Party Problem like situations. The results showed that adding these features improved the model’s ability to predict neural responses, with the correlation between predicted and recorded EEG increasing from 3.2% when using only the envelope to 3.8% when combining the envelope, phonemes, and Whisper’s layer-wise outputs (p = 0.0021). When distinguishing neural responses between attended and ignored speech, the inclusion of these additional features did not influence the classification accuracy significantly compared to models using only basic acoustic features. This work has potential applications in hearing technology, enabling the assessment of hearing and the effects of various hearing devices. Furthermore, it could inform signal processing algorithms in hearing aids, enhancing speech comprehension tracking and improving auditory attention by tailoring hearing devices to the specific needs of individual users.