Popular music lyrics are among the most widely circulated forms of performed language, influencing emotions but also reflecting societal values, voicing issues such as inequality and discrimination. This project investigated the link between our moral values and musical choices through lyrics and audio analyses. It developed novel NLP tools that address the challenge of automatically and rapidly detecting and characterising moral rhetoric in large music lyrics corpora.

Our theoretical framework for analysing morality is the Moral Foundations Theory, a well-established taxonomy that classifies moral evaluation along a small number of core moral traits, or foundations. This taxonomy is empirically grounded in social and cultural psychology and uniquely characterised by well-developed term dictionaries, lending itself to computational operationalisation in automated tagging and classification tasks

We first explored predicting moral values of listeners from self-reported genre-based preferences (CMMR 2021). We then used Facebook music-related Likes to extrapolate preferred artists and then predicted user moral values from lyrical and audio features of songs from preferred artists (ISMIR 2022, PLOS ONE 2023). Results showed that while genres provide insights into moral values comparable to demographics, lyrics and audio features can offer a more detailed understanding.

Based on the premise that moral values can be effectively captured in written narratives, we next introduced MoralBERT, a fine-tuned transformer-based language model (BERT) designed to detect moral foundations in social discourse data (ACM GoodIT 2024). The proposed model was subsequently adapted to tag music lyrics by further fine-tuning it on artificial lyrics generated by a large language model (GPT-4) and tested on a new annotated dataset of 200 real songs (ISMIR 2024).

Finally, using MoralBERT, we asked whether the moral expression of Western, English-language song lyrics shifted across six decades (Scientific Reports 2026). We combined 377,812 songs (1960–2010) from the WASABI dataset with 5,580 Billboard Year-End charting songs (1960–2023). We used generalised additive models to track lyrical-moral trajectories. We also related moral expression to non-moral themes and sentiment. We found that vice-oriented foundations rose substantially across both corpora, while Care and Purity declined, alongside increasing negative sentiment, anger and disgust. These trajectories appeared to differ by artist gender and genre.

Computational experiments with the 200 morally-annotated English song lyrics showed that certain timbral, harmonic and melodic features align with specific lyrical moral themes (CBMI 2025). Songs where lyrics express care or kindness featured clearer melodic lines and smoother, more harmonically stable timbral structures. In contrast, lyrics associated with harm or aggression were more often accompanied by rougher or noisier timbres and wider intervallic leaps (pitch or note ranges), the latter often strongly associated with high arousal emotions. Importantly, these associations could be identified from the sound of the music alone: lyrics were annotated in isolation, without the accompanying music, and machine learning models were then trained only on the audio content of the songs to predict their lyrics annotations.

Viewed together with work on the gendering of music composed for toy commercials, these findings reveal recurring cultural patterns between how music is used to encode moral meaning and how it is used to encode gender:

• musical consonance and smoothness signal empathy and femininity;
• dissonance and roughness signal aggression and masculinity.

Music thus becomes a shared expressive medium through which different worldviews are mapped onto the same psychoacoustical cues. Music is therefore more than a cultural and creative pursuit: it is also a social practice, a form of civic but also moral infrastructure, which raises urgent questions about social and moral progress in an era of increasingly technology-mediated music listening and engagement.