Spectral Feature Based Speech Recognition for Indigenous Chakma Language Using Deep Learning

Speech recognition is a remarkable process in the realm of human-computer interfaces that enables communication with machines, and deep learning has emerged as the most reliable approach for speech recognition. However, when it comes to working with tribal languages like Chakma, which lack abundant resources such as aligned speech and text, significant challenges arise. Despite some initiatives by the Bangladesh government to archive tribal languages, there has been limited technological work focused on the indigenous languages spoken in Bangladesh. This research aims to address this untapped sector by developing a spectral feature-based audio classification system using deep learning approach specifically tailored for Chakma words. The primary focus is to contribute to the field of Chakma language processing. The proposed system using VGG16 model achieves 95% accuracy. Additionally, a web-based audio data collection tool and an augmentation application have been developed to facilitate the visualization and generation of augmented audio and image data. This study not only emphasizes the potential of deep learning in handling tribal languages but also offers a foundation for further technological advancements in indigenous language processing.