Trainable data manipulation with unobserved instruments

Machine learning algorithms are the core components in a wide range of intelligent music production systems. As training data for these tasks is relatively sparse, data augmentation is often used to generate additional training data by slightly altering existing training data. User-defined techniques require a long parameter tuning process and typically use a single set of global variables. To address this, a trainable data manipulation system, termed player vs transcriber, was proposed for the task of automatic drum transcription. This paper expands the player vs transcriber model by allowing unobserved instruments to also be manipulated within the data augmentation and sample addition stages. Results from two evaluations demonstrate that this improves performance and suggests that trainable data manipulation could benefit additional intelligent music production tasks