The article discusses a recurrent model designed for audio-sheet music retrieval. The model learns joint embeddings to summarize longer passages of corresponding audio and sheet music. This method only requires weakly aligned audio-sheet music pairs and handles tempo variations. Experiments on synthetic and real piano data show that this method leads to more accurate retrieval in all configurations.
Publication date: 25 Sep 2023
Project Page: https://jku.at
Paper: https://arxiv.org/pdf/2309.12111
