AImoclips: Benchmarking Emotion Conveyance in Text-to-Music Generation Using a Dimensional Valence-Arousal Framework

Why It Matters

The benchmark uncovers systematic biases in current TTM systems, guiding developers toward more reliable emotional expression. Accurate emotion conveyance is essential for use cases such as film scoring, advertising, and therapeutic music, making these insights critical for the AI‑generated content market.

Key Takeaways

• AImoclips benchmark tests emotion conveyance in text-to-music generation.
• Six TTM systems evaluated with 991 instrumental clips and 12 emotion prompts.
• Human listeners rated valence and arousal, achieving above‑chance quadrant accuracy.
• Commercial models skew positive; open‑source models skew negative in valence.
• Higher audio quality (lower FAD) links to stronger emotion perception.

Pulse Analysis

The rapid rise of text‑to‑music (TTM) models has opened new creative avenues for marketers, filmmakers, and game developers, yet the industry has lacked a rigorous way to gauge whether these AI‑crafted tracks truly reflect the emotional intent of a prompt. Traditional music evaluation focused on timbre or genre similarity, ignoring the nuanced affective cues that listeners rely on. By framing emotion in the well‑established valence‑arousal space, AImoclips provides a quantifiable metric that aligns with psychological research and offers a common language for developers and stakeholders.

AImoclips’ methodology combines a sizable dataset—991 instrumental clips generated from six distinct TTM systems—with a large‑scale human study involving 111 participants who supplied 6,162 valence and arousal ratings on a nine‑point scale. The findings reveal that while all models outperform random guessing at the quadrant level, their overall precision is limited. Notably, commercial platforms tend to produce music with a positive valence bias, whereas open‑source alternatives skew toward negative valence, each exhibiting distinct arousal patterns. Moreover, the study links higher audio fidelity, measured by Fréchet Audio Distance, to stronger emotional perception, and shows that CLAP‑based text‑audio alignment primarily predicts valence outcomes.

These insights have immediate implications for product roadmaps and investment decisions. Companies aiming to integrate AI‑generated music into emotionally driven experiences—such as adaptive game soundtracks or personalized wellness playlists—must prioritize improvements in both audio quality and semantic alignment to reduce bias. The benchmark also sets a research agenda, encouraging the development of models that can reliably map nuanced linguistic cues to musical affect, ultimately expanding the commercial viability of AI‑composed soundscapes.