Evaluating the generalization ability of deep learning models for sound source localization

Эндаумент фонд ТГУ!
Впиши своё имя в историю университета
- Сделать пожертвование -

Add to "Quick Collection"

Description	Size	Format
Evaluating the generalization ability of deep learning models for sound source localization	497 KB	Adobe Acrobat PDF	Read	Download
DOI Доступ к ресурсу на сайте издателя			10.17223/19988605/72/11

Title: Evaluating the generalization ability of deep learning models for sound source localization
Title: Оценка обобщающей способности моделей глубокого обучения для локализации источника звука
Creator: Shahoud, Ghiath M.
|
Agafonov, Evgeny D.
Date: 2025
DOI: 10.17223/19988605/72/11
Description: In this paper, the generalization ability of deep learning models used to solve the sound source localization problem with a spatial resolution of 10º is evaluated when the configuration settings are changed. The generalization ability of the models was evaluated in a closed reverberant environment using an orthogonal microphone array. Two models were considered: SI-GCC-CNN, which is based on combining the features of sound intensity and generalized cross-correlation - phase transform as input data for convolutional neural networks, and SI-CNN, which is based on feeding the features of the sound intensity into the convolutional neural network. Simulation and modeling results show that the SI-GCC-CNN model is effective in its generalization ability and outperforms the SI-CNN model, achieving an improvement in localization accuracy by 22,1% when changing the size of the room, by 15,6% when changing the location of the microphone array and by 32% when changing the distance between the source and the center of the microphone array.