Automatic language identification (LID) systems typically perform well when dealing with high-resource languages, but their performance declines with low-resource languages. Moreover, when training data consists mainly of clean, read speech and test data includes real-world situations such as spontaneous speech from noisy or uncontrolled environments, the system’s performance suffers due to domain shift. This paper addresses the challenge of domain shift, focusing on 12 low-resource Indian languages, many of which belong to the same language family. We propose two frameworks that integrate phonotactic features with advanced deep learning frameworks to enhance LID performance under domain shift conditions. By leveraging self-supervised learning techniques, we extract phoneme posterior probabilities, improving the representation of acoustic features beyond low-level features. Our experiments demonstrate the effectiveness of this approach.