This research presents an approach to automatic melody generation from lyrics by leveraging deep learning models on a large dataset of symbolic music data. The dataset includes a vast collection of MIDI files, each providing rich musical details such as lyric syllables, corresponding notes, key, tempo, number of instruments, and track length. This data is pre-processed to align musical notes with syllables, creating a structured basis for melody generation. BERT embeddings are applied to lyrics to capture the global context and feel of each song, ensuring that semantic nuances at the syllable level align with the overall mood and style. Each syllable is represented with a corresponding musical note chosen from a diverse set of pitch possibilities. These embeddings, combined with note representations, are used to generate song data sequences which are then fed into LSTM, BiLSTM, and GRU networks for melody prediction. Model performance is evaluated using pitch distribution metrics and statistical measures like mean and median, mode, range, standard deviation. By comparing predicted note sequences with actual note sequences, this framework facilitates cross-modal learning between text and music, establishing a foundation for advanced applications in AI-driven music composition.