Classification of Harmful Noise Signals for Hearing Aid Applications using Spectrogram Images and Convolutional Neural Networks

Hearing aids are important for impaired people suffering from hearing loss. Generally, digital hearing aids are preferred since digital systems are flexible and programmable for different hearing loss conditions. Hearing aid applications usually use low-pass filters, adaptive filters and spectral analysis to remove harmful noise and amplify incoming speech signals. However, existing methods in digital hearing aids may not remove all harmful noises. To overcome this problem, first there is a need to understand whether the speech signal is harmful or harmless. If the signal is harmful, noise type should be identified, and then according to the noise type the filter will be selected. In this paper, we particularly focus on classification of speech signals into different classes. One of these classes is the clean speech signal class, and the others are harmful noisy speech signal classes that are generated with respect to five different noise types. We propose to use spectrogram images of incoming speech signals and a Convolution Neural Networks (CNN) for accurate classification. We create a dataset that consists of speech signals corrupted by different noise types such as white noise, jet aircrafts noise, storm noise, running tape noise and high-frequency noise. We compare the performance of different CNN architectures, and also conduct evaluation for computation time. Combination of spectrogram images with CNNs is a new approach for harmful speech classification. Results show that the proposed method can classify speech signals very accurately for different noise types.