Acoustic Scene Classification in Hearing Aids Noise Suppression Applications
Published in Master Thesis, 2025
In this work, a study of Acoustic Scene Classification (ASC) for hearing aids was conducted, leading to the development of an industrial ASC model based on a convolutional neural network designed to run in hearing aids for noise suppression purposes. All phases of an ASC project were analyzed, starting with dataset construction and data labeling, where an auto-labeling method was designed to streamline the labeling process. Significant emphasis was placed on miniaturizing the model, evolving from a baseline model of 783 MFLOPs based on audio embeddings using VGGish to a production model running at just 1.6 MFLOPs. The technological approaches used for model miniaturization included knowledge distillation, dimensionality reduction, and DepthWise convolutions.
Recommended citation:
Download Paper | Download Slides | Download Bibtex