Audio synchronization aims at aligning multiple recordings of the same piece of music. Traditional synchronization approaches are often based on dynamic time warping using chroma features as an input representation. Previous work has shown how one can integrate onset cues into this pipeline for improving the alignment's temporal accuracy. Furthermore, recent work based on deep neural networks has led to significant improvements for learning onset, beat, and downbeat activation functions. However, for music with soft onsets and abrupt tempo changes, these functions may be unreliable, leading to unstable results. As the main contribution of this paper, we introduce a combined approach that integrates activation functions into the synchronization pipeline. We show that this approach improves the temporal accuracy thanks to the activation cues while inheriting the robustness of the traditional synchronization approach. Conducting experiments based on string quartet recordings, we evaluate our combined approach where we transfer measure annotations from a reference recording to a target recording.

Direct link to video