Communication serves as a fundamental link among individuals globally; without this medium of engagement, people risk losing their sense of purpose in life. In particular, communication poses a challenge for hearing individuals and those who are deaf or unable to hear. Many existing systems and applications have demonstrated their ability to recognize the sign language of deaf or mute people. However, these applications are still useful in standard or basic situations but not in critical scenarios, such as at night. This type of case remains underexplored and requires critical consideration. Although many studies based on images are affected by the luminosity of the captors, the sensitivity of gesture images requires a high level of preprocessing before they can be classified, and certain significant features may disappear after deep treatment. In this context, we propose a study to build an application that can classify gestures under various conditions, such as at night or in the dark. As a second objective, we needed to validate whether the variation in data affected the accuracy of gesture recognition. This paper presents a novel method for addressing limitations related to low-quality resolution images by recognizing five ASL gestures in the dark using raw data from FMCW radar and a convolutional neural network (CNN) after collecting them in various orientations or angles to analyze the impact of dataset variability on the model.