Some people cannot produce sound although their facial muscles work properly due to having problem in their vocal cords. Therefore, recognition of alphabets as well as sentences uttered by these voiceless people is a complex task. This paper proposes a novel method to solve this problem using non-invasive surface Electromyogram (sEMG). Firstly, eleven Bangla vowels are pronounced and sEMG signals are recorded at the same time. Different features are extracted and mRMR feature selection algorithm is then applied to select prominent feature subset from the large feature vector. After that, these prominent features subset is applied in the Artificial Neural Network for vowel classification. This novel Bangla vowel classification method can offer a significant contribution in voice synthesis as well as in speech communication. The result of this experiment shows an overall accuracy of 82.3 % with fewer features compared to other studies in different languages.