Penanganan Noise pada Model Klasifikasi Perintah Suara Menggunakan Complete EEMD with Adaptive Noise (CEEMDAN)

Abstract

The application of voice recognition systems is currently widely used in various fields, one of which is voice commands classification. Voice classification is basically done by using feature extraction to recognize the characteristics of each signal. The MFCC method is used in this study. MFCC is a feature extraction method that calculates the cepstral coefficient by imitating the human hearing systems. In this study, the LSTM model with a bidirectional scheme (Bi-LSTM) is used to build a voice command classification model. This method is considered to be able to accommodate the characteristics of voice commands which have a short duration and sensitive to noise. In order for the classification model to maintain high accuracy, CEEMDAN is used as a noise reduction method. CEEMDAN is considered capable of decomposing noisy signals and separating the original signal from its noise component. The results showed that the Bi-LSTM model had an accuracy of 96% and an F1-score of 95%, but experienced a significant decrease on performance when given by 10 Signal to Noise Ratio (SNR) to 0 SNR noise level. In this study, CEEMDAN has successfully improve the performance of the model on noised data. CEEMDAN is able to increase the accuracy and F1-score by an average of 29.57% and 33.43%, which makes the classification model more robust in classifying voice commands at certain noise levels. The results of the comparison with other noise handling methods show that CEEMDAN performs better in noise reduction.