*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title:
Extendable Learning for Automatic Music Transcription —
Improving Automatic Drum Transcription with Labeled and Unlabeled Data
Abstract:
Automatic drum transcription is a sub-task of Automatic Music Transcription (AMT) that involves the conversion of drum-related audio events, such as drum onset times and playing techniques, into musical notations. While noticeable progress has been made in the past by combining the pattern recognition methods with audio signal processing techniques, the major limitation of many state-of-the-art systems still originate from the difficulty of obtaining a meaningful amount of labeled data to support the data-driven algorithms. To address the challenge of insufficient labeled data, this work attempts to explore the possibility of utilizing unlabeled music data from online resources. Specifically, this work focuses on improving the drum transcription systems with two proposed methods. The first method is based on the concept of distillation, which improves the performance of a model by creating an ensemble system followed by a model compression process with the unlabeled data. The second method is based on the concept of self-taught learning, which learns appropriate feature representations from unlabeled data in order to achieve improvements. The contributions of this work will be: first, it will provide insights into the viability of leveraging easily accessible unlabeled music data for drum transcription. Second, it will enable a scheme for data-driven algorithms to benefit from the unlimited online resources and thus might have impact on many other audio and music related analysis tasks traditionally impeded by small amounts of labeled data. Last but not least, the resulting system could improve the state-of-the-art in automatic drum transcription
Committee members:
Dr. Alexander Lerch
Dr. Jason Freeman
Dr. Mark Clements (ECE)
