SEMINAR:Source Separation of Piano Music Recordings

Guest: Yiğitcan Özer, Research Intern at SONY AI, Tokyo

Title: Source Separation of Piano Music Recordings

Date/Time: October 9, 2024, 13:40

Location: https://sabanciuniv.zoom.us/j/94589085437,

Meeting ID: 945 8908 5437

Abstract: This work addresses the novel source separation task of decomposing piano concerto recordings into individual piano and orchestral tracks, enabling pianists at all levels to practice and perform with orchestral accompaniments. As one main contribution, we adapt deep-learning-based source separation techniques, initially designed for the separation of popular music recordings or speech signals. In particular, we address the challenge of higher spectro–temporal correlations between piano and orchestra compared to popular music tracks, and the lack of multitrack datasets for training, by introducing musically motivated data augmentation approaches. Another main contribution of this research is the creation of a multitrack dataset of piano concertos, which encompasses a collection of excerpts with separate orchestral and piano tracks, performed by both professional and amateur pianists. The broad scope of our dataset not only serves as a valuable resource for both quantitative and subjective evaluation of source separation models but also opens up various possibilities for other music processing applications, including score following, downbeat estimation, and music synchronization. As a third main contribution, we split the separated piano tracks into notewise events using score-informed nonnegative matrix factorization (NMF). In particular, we apply this audio decomposition technique for evaluating source separation results of piano tracks introducing a notewise signal-to distortion ratio (SDR) measure to gain deeper insights into various source separation artifacts. Overall, this work not only addresses a novel challenge in the field of music information retrieval but also enhances the way pianists can interact with classical music performances.

Bio: Yigitcan Özer received his B.Sc. degree in Electrical and Electronics Engineering from Bilkent University, Ankara, Turkey, in 2013, and his M.Sc. degree in Communications Engineering from the Technical University of Munich, Germany, in 2015. He completed his Ph.D. at the International Audio Laboratories Erlangen, Erlangen, Germany, under the supervision of Prof. Meinard Müller, where he also served as a graduate research assistant from January 2021 to August 2024. Prior to his Ph.D. studies, he worked as a Research Associate in the Spoken Language Processing Group at the Fraunhofer Institute for Integrated Circuits IIS, Erlangen. He is currently a research intern at Sony AI, Tokyo. His research interests include audio source separation, music synchronization, text-to-speech synthesis, and audio watermarking.