This study developed an Assamese parts-of-speech (POS) tagger employing advanced neural network architectures such as Long Short-Term Memory with Convolutional Neural Networks (LSTM + CNN) and Bidirectional Long Short-Term Memory with Convolutional Neural Networks (Bi-LSTM + CNN). The major goal was to increase the accuracy of Assamese POS tagging using the Bureau of Indian Standards (BIS) tag set. A huge Assamese text corpus was annotated using BIS rules, and it went through significant preprocessing to ensure cleanliness and effective tokenization. The data was translated into numerical representations suitable for model training. The design consisted of an embedding layer with pre-trained word embedding to capture linguistic nuances, LSTM and Bi-LSTM layers to learn sequential dependencies, CNN layers to extract local patterns. Experiments with various configurations found that the LSTM + CNN model reached 87.3% accuracy, whereas the Bi-LSTM + CNN model achieved a substantially better accuracy of 91.51%. These findings reflect a considerable improvement in Assamese POS tagging, with implications for linguistic study and applicability to other languages.