Data-Driven Analysis of Text-Conditioned AI-Generated Music: Insights from Suno and Udio

AI-generated music platforms such as Suno and Udio let users create songs from text prompts, and these tools now serve hundreds of thousands of users. AI music has begun to appear in advertising and on music charts in several countries, and professional creators are integrating these tools into their workflows. This study asks how these platforms are used and what subjects inspire their users. By analyzing a large, data-driven corpus of user-generated songs, the paper moves beyond the in-depth study of individual works toward large-scale patterns in an emerging cultural practice.

No prior work has systematically studied prompting in music generation, so the study draws on research in text-to-image prompt analysis, such as work on DiffusionDB and Midjourney data that used HDBSCAN and UMAP to build taxonomies of prompt specifiers. In MIR, natural language processing has been applied to lyrics for genre and mood classification, similarity measurement, and sociocultural analysis, including studies showing changes in lyrical complexity, emotion, and sentiment across decades of popular music.

Both platforms center on a text prompt box. Suno offers a Custom mode for direct control over lyrics and style tags and an instrumental option, while Udio translates a general prompt into tags unless Manual Mode is enabled, and supports metatags in square brackets added to lyrics. Both accept free text, including emojis and non-Latin characters. Metadata fields are inconsistently defined across platforms: Suno's 'prompt' refers to lyrics input, while Udio's 'prompt' is the user input. The authors harmonize these fields in their analysis. Post-processing also differs, as Udio replaces real artist names with generic tags, whereas Suno blocks them.

The method has six steps: scraping the public 'new songs' playlists; extracting prompt, lyrics, and tag metadata into dataframes; embedding text with a pretrained model (NV-Embed-v2, with 4096-dimensional representations); reducing dimensionality with UMAP; clustering with HDBSCAN, which automatically detects outliers; and automatically naming clusters via TF-IDF over the top three lemmatized terms, followed by manual refinement and a 2D scatter visualization. Non-English entries are filtered with fastText language identification to avoid issues with English-tuned NLP tools, and tags are mapped to macro categories such as genres, instruments, and technical terms like BPM. Hyperparameters were tuned by trial and error.

The dataset contains 101,953 songs, of which 20,519 come from Udio and 81,434 from Suno, collected between May and October 2024. Key fields include id, title, tags, lyrics, prompt, and optimized or description prompts. Because field semantics differ between platforms, the authors define unified terminology: 'lyrics' (Suno prompt / Udio lyrics), 'tags' (tags and their post-processed versions), and 'prompt' (Suno gpt_description_prompt / Udio prompt). The authors also share code and a list of song URLs to support reproducibility, and the SONICS dataset was noted as a separate resource for AI music detection.

English dominates lyrics on both platforms, at 71.39% on Udio and 46.75% on Suno. Other prevalent languages include German (3.68% Udio, 8.87% Suno), Russian (2.99% Udio, 6.68% Suno), Spanish, Portuguese, Korean, Chinese, Indonesian, French, Japanese, Turkish, Italian, Thai, Vietnamese, and Polish. The top 15 languages account for over 90% of the dataset. The distribution reveals notable differences in language preference between the two platforms, with Suno showing considerably more non-English usage.

Clustering of prompts and tags reveals emerging prompting strategies and the qualifiers users most often add to style prompts, such as genre, instrumentation, tempo, and mood vocabulary. Suno's free-text style field often produces long individual tags, while Udio's tag post-processing normalizes entries. These patterns show how users translate musical intent into textual conditioning, revealing shared conventions that resemble the prompt engineering practices seen in image generation.

Clustering lyrics surfaces the most common themes and shows how many prompts reference real musicians or attempt to impersonate artists. Users also experiment with metatags in square brackets to steer song structure and performance, a peculiar form of model steering. These findings highlight both creative expression and the boundary-testing behaviors that accompany a new, text-driven music practice.

The analysis suggests that AI music platforms are shaping new cultural practices around prompting, language, and lyrical expression. The methodology can be extended to other platforms and time periods, and comparative studies could examine how interfaces and post-processing policies influence user behavior. Future work includes deeper musicological analysis of audio, study of community discourse, and tracking how these practices evolve as platform features change.

This study provides a systematic, data-driven view of how users engage with text-to-music AI platforms Suno and Udio. By combining state-of-the-art text embeddings, dimensionality reduction, and clustering, it reveals prevalent languages, prompting strategies, lyrical themes, and metatag steering. The authors contribute a reproducible methodology, shared resources, and interactive visualizations to support the emerging field of AI music studies.