Beyond pitch: enhancing sight-singing evaluation with solfege
Hongfeng Gao, Wei Hu · IET conference proceedings. · 2025
We present an innovative automatic sight-singing evaluation system that significantly improves notetranscriptionandalignmentaccuracy. Our system enhances traditional Automatic Singing Transcription (AST) by incorporatingsolfege(e.g., do, re, mi) predictions during transcription. We use wav2vec2 as the backbone network and introduce aU-Netmodule withconsistency lossto improve note boundary detection, while ensuring consistency between pitch, solfege, and boundary predictions. For alignment, our system leverages solfege information instead of traditional pitch-based methods, achieving state-of-the-art accuracy and handling common beginner errors such as pitch deviations. Experiments show anF1-score of 85.91%on the SSVD dataset, setting new benchmarks. Our AST model also achieves SOTA results in pure AST tasks, validating its design. This combination of advanced AST and solfege-based alignment makes our system robust and reliable for real-world sight-singing evaluation.