TLDR: This research paper analyzes over 100,000 AI-generated songs from Suno and Udio (May-Oct 2024) to understand user behavior. It reveals diverse prompting strategies, prevalent lyrical themes like love, worship, daily life, and fiction, and how users employ tags and metatags to steer music generation. The study also highlights differences in how platforms handle artist names and the global linguistic diversity of users, offering a foundational understanding of the emerging AI music landscape.
The world of music creation is undergoing a significant transformation with the rise of Artificial Intelligence (AI) platforms like Suno and Udio. These platforms allow users to generate music from simple text prompts, and their popularity has soared, with hundreds of thousands of users creating songs that are even appearing in advertising and charting in various countries. A recent research paper delves into how these platforms are being used and what inspires their users, offering a fascinating glimpse into the evolving landscape of AI-generated music.
Titled “Data-Driven Analysis of Text-Conditioned AI-Generated Music: A Case Study with Suno and Udio,” the study was conducted by Luca Casini, Laura Cros Vila, David Dalmazzo, Anna-Kaisa Kaila, and Bob L.T. Sturm from KTH Royal Institute of Technology in Stockholm, Sweden. Their work provides the first systematic analysis of user behavior on these platforms, utilizing a large dataset of over 100,000 songs generated between May and October 2024.
Unpacking the Data: How the Study Was Conducted
To understand user patterns, the researchers collected textual metadata including prompts, tags, and lyrics from both Suno and Udio. They employed a sophisticated data-driven methodology, borrowing techniques from natural language processing and data science. This involved using state-of-the-art text embedding models like NV-Embedv2, followed by dimensionality reduction (UMAP) and clustering methods (HDBSCAN). These tools helped them to automatically annotate and visualize the processed data, revealing prominent themes and strategies.
What Users Are Creating: Key Findings
The analysis uncovered several intriguing trends:
Language Diversity: While English is the most prevalent language for lyrics, the platforms see a wide variety of other languages, including German, Russian, Spanish, Portuguese, Korean, and Chinese. This suggests a global user base, though the distribution doesn’t necessarily mirror the most commonly spoken languages worldwide.
Prompting Strategies: Users employ diverse prompting strategies. Some prefer a list of comma-separated qualifiers (e.g., “modern country, contemporary folk, introspective”), while others use more descriptive, literary prompts (e.g., “A song about…” or “A jazz ballad with trumpet…”). The study also identified “scripted” prompts, which follow a consistent pattern, possibly generated by external services. Interestingly, users sometimes publish multiple attempts at generating the same song, indicating an iterative creative process.
Themes in Lyrics: The lyrical content generated by users spans a vast array of themes. Love, in its many forms (unrequited, breakups, longing), is a dominant subject. Worship songs, particularly related to Christianity and Islam, form a significant cluster. Users also create music for celebrations like birthdays and holidays, or to commemorate daily routines, trips, and sports events, highlighting a “musicalization of everyday life.” Other popular themes include pets, animals, food, and even current geopolitical events like the 2024 USA elections and the Israeli-Palestinian conflict. Fiction, such as high fantasy, horror, and video games (like Pokémon and Helldivers), also inspires many songs. A thematic cluster around technology shows an equal number of songs praising and criticizing AI. Humorous or meme-related lyrics are also common, featuring popular online phrases.
The Role of Tags: Tags are crucial for conditioning the AI models on genre and style. Suno tends to have longer, more free-form tag descriptions, whereas Udio users often employ shorter, comma-separated descriptors. The researchers categorized tags into high-level groups like Genre/Style, Instrument, Qualifiers/Mood, and Voice, showing clear separation and clustering of related terms. For instance, guitar-related tags cluster near rock and country genres, while piano tags are closer to jazz.
Handling Artist Names: A notable difference between the platforms is how they handle real artist names. While Suno generally uses prompts only for lyrics and forbids artist names in tags, Udio is more flexible. Udio automatically extracts tags from prompts and, in many cases, allows artist names. However, it replaces these names with generic descriptors in the metadata, likely to avoid intellectual property issues. For example, “XXXTentacion” might be replaced with “emo rap, alternative r&b, hip hop.”
Metatags for Control: Users often embed “metatags” within lyrics, typically enclosed in square brackets (e.g., [verse], [chorus], [guitar solo]). These metatags serve as instructions to steer the music generation, influencing structure, instrumentation, delivery, and dynamics. While users experiment with elaborate instructions, very long sequences can sometimes be ignored or misinterpreted by the AI. Udio actively encourages this experimentation by providing tool-tips with suggestions. Interestingly, some Suno users attempt to evoke specific artist voices or styles using metatags, circumventing the platform’s restriction on artist names in tags.
Also Read:
- Navigating the AI-Generated Content Landscape: Trends, Challenges, and Future Directions
- Simulating Music Discovery Dialogues with Advanced AI
Looking Ahead: Limitations and Future Directions
The study acknowledges certain limitations, such as its six-month data collection period, which might not capture all evolving trends. The focus on publicly published songs also introduces a bias. The researchers emphasize the need for future work to explore non-English prompts and lyrics to uncover cultural differences in AI music engagement. They also suggest improving outlier detection in their clustering methodology and using large language models for more sophisticated cluster naming. Further ethnographic studies and integration of data from community forums like Discord and Reddit could provide deeper insights into user practices and the reception of AI music.
This pioneering research offers a comprehensive understanding of how users interact with AI music generation platforms. It highlights the diverse creative expressions and strategies emerging in this new cultural practice, paving the way for further exploration into the fascinating intersection of AI and music. You can read the full research paper here.
