LLM-Guided Affective Music Generation via Prompt-to-Parameter Mapping and Frequency-Aware Evaluation

Abstract

Recent text-to-music generation models have enabled users to compose music from natural language description; they typically regard emotional prompts as a black-box conditioning signal, which is hard to interpret, control, and assess. In this paper, we present an LLM-guided affective music creation framework featuring prompt-to-parameter transformation and frequency-aware evaluation. Our approach begins by taking a handwritten emotional cue and applying a large language model as affective mapper that converts subjective language into structured music parameters such as tempo, dynamics, harmony, texture density, register, instrumentation and spectral targets. These parameters are then input into a downstream model that generates music. For music generation evaluation, we propose a multi-dimensional protocol on the aspects of audio fidelity, text-music alignment, and frequency-domain consistency. We evaluate 40 affective prompts, and our method achieves 0.536 of CLAP alignment, 4.31 of FAD, and 0.792 of spectral consistency, while direct prompt generation achieves 0.412, 5.87, and 0.641, respectively. Subjective listening results also reveal that our method results in higher overall quality, emotional consistency and listening comfort. These results show that the explicit LLM-based affective mapping offers a more interpretable and controllable approach to emotional music generation.