PATER: Pedal-Aware Transformer for Symbolic Music Emotion Recognition

Abstract

Emotion-aware understanding of symbolic music is increasingly useful for digital entertainment applications such as game music organization, affect-aware retrieval, and interactive soundtrack systems. Symbolic Music Emotion Recognition has advanced with large-scale pre-training. Yet existing models often underrepresent expressive performance controls, especially sustain-pedal use, by treating them as simple binary tokens, thereby overlooking expressive cues relevant to emotion perception. We propose PATER, a pedal-aware transformer model for symbolic music emotion recognition that combines Teacher-Guided Pedal Standardization with a Pedal-FiLM conditioning block. By injecting teacher-derived pedal-state priors into symbolic note representations, PATER enables the model to incorporate expressive performance cues that are typically absent from note-only SMER pipelines. Because the evaluated SMER datasets do not provide reliable measured sustain-pedal annotations, PATER does not assume access to ground-truth pedaling. Instead, it uses a frozen SusPedal teacher to derive token-aligned pseudo-pedal state distributions over four functional states (Full, Half, Quick, and Off), which serve as musically structured conditioning priors. The results suggest that these teacher-derived priors can improve matched note-only baselines in the evaluated settings, although the VGMIDI results should be interpreted cautiously due to the dataset's limited size and unresolved Q3 collapse.