Preliminary Studies for Cracking Sound Identification During Coffee Roasting

A coffee roasting plays an important role in determining the taste of coffee. The process consists of drying, browning, first crack, roast development, and second crack. The first crack is a beginning to form the characteristics of coffee beans. At this stage, a professional coffee roaster must ensure the spesific temperature and time, hence the beans are not getting burn. In manual coffee roasters, this causes inconsistency, because it really depends on the ability and experience of the coffee bean roaster. Therefore, in this study, we conducted a machine learning-based preliminary study to detect the cracking sound of coffee beans. The preliminary study was conducted by collecting videos of the roasting process from the internet. The videos are converted into audio format and separated between audio containing the sound of coffee beans cracking and audio that does not contain the sound of coffee beans cracking. The audios are then converted into an image format using the Mel-frequency cepstral coefficients (MFCC) method and modeled using convolutional neural network (CNN). From the 216 dataset files, we obtained an accuracy value of 89.7% with data lost during the testing process as much as 34.9% with 50 epochs of training and validation processes.