In recent years, the silent killer disease, defined as a non-communicable disease, has become a frequent topic discussed in many academic discussions. Although this disease is not transferable from one to another, starting from 1990, the increment trend was annually published by the world statistic data for this disease, e.g., heart attack and stroke. The more significant consequence of these two diseases is to disable one or more human capabilities. One of the stroke disease effects is becoming disabled from hearing. Speech disabilities are the focus of this proposed study in this paper. Since the person diagnosed as a stroke patient requires attending the recovery session or rehabilitation session, the rehabilitation center must prepare and provide a sound module and system to help the patient regain their capability. Rehabilitation is an alternative path to gradually giving routine practice to the patient to improve their capability back. For this purpose, the rehab center requires a quantity of time to provide the patient to attend the training session. The training, however, is conducted in two ways, physically and virtually. For the Malaysia stroke patient, the training for pronouncing the vowel in the Malay language is crucial in getting back the speaking capability. Since the Malay language has 6 types of vowels, which are /a/, /e/, / ^/, /i/, /u/, and /o/. Here, there is a limitation to smartly recognizing the difference between the two /e/ vowels. Malay's /e/ vowel is crucial as the similar spelling vocabulary conveys two different meanings. This study analyzed the differences in recognizing the two /e/ vowels using Convolution Neural Network (CNN) with the help of the existing sound-image dataset.