DEEP DENOISING BABBLE NOISE USING A COMBINATION OF CNN AND RNN

• Main Author: Nurul Hukmi, Imam
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/76866
• Institution: Institut Teknologi Bandung

Babble noise or chattering noise has become a type of sound signal noise that is difficult to handle. In this study, this noise handling was carried out by utilizing a deep learning model, more precisely a combination of CNN and RNN-based architectures. The model with this basis was chosen because of the consideration of the advantages possessed by each type. CNN-based architectures can handle spatial data, have better generalization, and can do inference quickly. RNN-based architectures are specialists in dealing with ordered data. The built model is trained to be able to identify the noise contained in the noise signal in the form of a spectrogram. To find out how good the performance of the model is, this blended model is compared with CNN-based models and RNN-based models. Especially for models based on a combination of CNN and RNN as the main model, experiments will be carried out to get the best configuration of the model architecture. The metric that will be used is PESQ (perceptual evaluation of speech quality), which is a metric that is built to resemble how humans evaluate signal quality and assess the difference between the noise signal and the reference clean signal. The experimental results show that the best configuration for the CNN and RNN combination model is the model with the CNN U-Net architecture, the activation function is PRELU, and the RNN layer type is GRU. After conducting the training, it was found that the model based on the combination of CNN-RNN, CNN, and RNN had PESQ values of 2.23, 1.95, and 1.58, respectively. The results of the follow-up research show that the blended model succeeded in providing better output than the other two models on data with SNR levels within its training range.