Abusive music and song transformation using GenAI and LLMs

TL;DR

This study tackles the problem of abusive content in music by applying GenAI and LLMs to transform both lyrics and vocal delivery while preserving melody. The authors integrate a pipeline that uses Demucs for vocal separation, GPT-4 for lyric rewriting, and Suno AI for vocal synthesis to produce safer-transformed tracks, evaluated through lyric sentiment (DistilBERT), semantic similarity (MPNet), and acoustic metrics ($HNR$, $CPP$, $Jitter$, $Shimmer$). They report substantial reductions in aggression (63.3%–85.6%) and improvements in vocal quality across four songs, indicating the potential of transformation-based content moderation as a safer listening approach. Limitations include a small sample size, variability in AI outputs, and beat alignment issues, with future work proposed to scale up, conduct perceptual studies, and explore real-time implementations for streaming platforms.

Abstract

Repeated exposure to violence and abusive content in music and song content can influence listeners' emotions and behaviours, potentially normalising aggression or reinforcing harmful stereotypes. In this study, we explore the use of generative artificial intelligence (GenAI) and Large Language Models (LLMs) to automatically transform abusive words (vocal delivery) and lyrical content in popular music. Rather than simply muting or replacing a single word, our approach transforms the tone, intensity, and sentiment, thus not altering just the lyrics, but how it is expressed. We present a comparative analysis of four selected English songs and their transformed counterparts, evaluating changes through both acoustic and sentiment-based lenses. Our findings indicate that Gen-AI significantly reduces vocal aggressiveness, with acoustic analysis showing improvements in Harmonic to Noise Ratio, Cepstral Peak Prominence, and Shimmer. Sentiment analysis reduced aggression by 63.3-85.6\% across artists, with major improvements in chorus sections (up to 88.6\% reduction). The transformed versions maintained musical coherence while mitigating harmful content, offering a promising alternative to traditional content moderation that avoids triggering the "forbidden fruit" effect, where the censored content becomes more appealing simply because it is restricted. This approach demonstrates the potential for GenAI to create safer listening experiences while preserving artistic expression.

Figures (11)

• Figure 1: Music Transformation Framework showing stages including audio data collection and pre-processing, waveform and spectrogram analysis. We then implement vocal quality measurements and use the Suno AI to replace vocals, transforming the song.
• Figure 2: Music Evaluation Framework showing major stages that include collecting lyrical data, pre-processing data and sentiment analysis. The sentiment analysis pre and post transformation of the lyrics evaluates if sentiments are maintained in transformed songs. We also calculate the cosine similarity between the original and transformed songs for semantic analysis.
• Figure 3: Line graph of cosine similarity between the original and transformed versions of text with a rolling window size of 5.
• Figure 4: N-gram analysis of song lyrics showing the most frequent word sequences for each song.
• Figure 5: Waveform analysis separated by section