Adaptive distance normalization for real-time music tracking

The goal of real-time music tracking is to follow a musical performance on-line and at any time report the current position in the score. To achieve this, both the score and the performance have to be represented in a suitable way. In this paper, we first evaluate the performance of some well-known features and then propose a simple but effective distance normalization strategy for onset-emphasized features, which greatly improves the alignment results. Finally, we combine both harmonic and onset-emphasized features in a fashion known from off-line audio alignment, resulting in a combination which outperforms each individual feature regarding robustness and accuracy.