Malay words and dialect identification using long short-term memory and convolutional neural networks on trained Mel frequency cepstral coefficient / Mohd Azman Hanif Sulaiman

As Malaysia moves towards to the Industrial Revolution (IR 4.0), and as machines become more intelligent and autonomous, man and machine interaction are becoming inevitable. In general, the machine robustness towards dialect identification will be the main one of the many practical methods for interaction is using spoken language. However, there are many limitations of this type of interaction, particularly for native speakers other than English among them is dialect criteria for the system development. The complexity of dialects requires a new paradigm/generation of Artificial Intelligence (AI) - based classifiers methods capable of adapting to the linguistic of the language. These proposed techniques are recent and relatively unexplored in the field of dialect identification. This research explores two types of methods for dialect identification, namely Convolution Neural Network (CNN) for Malay dialect identification and MFCC feature extraction technique will be used to extract the features. Next, these features will be transferred to the CNN to be trained. For Long short-term Memory (LSTM), the inputs are fed directly from the recorded dataset for training. This research is to design and implement the CNN and LSTM network for Malay language dialect classification. In support of this, several objectives need to be achieved is to perform features extraction using MFCC on Malaysian Dialect, then to classify the Malaysian Dialect using CNN and LSTM neural networks and compare the performance of CNN and LSTM neural networks on Malay dialect identification.