Safet AIsović: Comparison of Methods for Generating Sevdah Music Lyrics

This paper explores the intersection of natural language processing (NLP) and traditional Bosnian Sevdah music, focusing on the generation of authentic Sevdah lyrics through advanced computational techniques. Using a comprehensive collection of Sevdah lyrics obtained through internet scraping, we employ two distinct LSTM-based approaches for text generation and fine-tuning of the pre-trained “BERTić” model. Our investigation explores cultural and linguistic elements present in Sevdah lyrics, aiming to capture the emotional depth and poetic essence into several models capable of generating new and unique Sevdah lyrics that can be later used as a basis or inspiration for Sevdah lyrics authors. Through a comparative analysis, we demonstrate that the second LSTM approach, implemented using TensorFlow and Keras, outperforms the alternative methods proposed in this paper. This research not only contributes to the emerging field of AI-driven lyrics generation, but it is also a pioneer project in the field of Sevdah music, aiming to preserve and promote cultural heritage through innovative applications of NLP technology. In summary, the study navigates the landscape of Sevdah lyric generation, comparing BERTić and two different LSTM approaches. The findings underscore the efficacy of these models, particularly the second LSTM approach, in generating Sevdah lyrics that encapsulate the rich cultural and emotional tapestry of this traditional Bosnian music genre.