Under CopyrightWeber, PhilippPhilippWeberUhle, ChristianChristianUhleMüller, MeinardMeinardMüllerLang, MatthiasMatthiasLang2024-10-222024-10-222024-09-30https://publica.fraunhofer.de/handle/publica/477872https://doi.org/10.24406/h-47787210.1109/IS262782.2024.1070413010.24406/h-477872This paper proposes the application of dynamic few-shot learning for automatic drum transcription (ADT). The contributions of this work are threefold. First, we adapt dynamic few-shot learning to improve the classification of superimposed events. Secondly, we introduce a novel method for generating training data for ADT. Thirdly, we demonstrate how our model can be applied in real-time without strongly deteriorating the classification performance. We evaluate transcription performance in the presence of melodic instruments for 10 drum classes on three publicly available test datasets and achieve state-of-theart performance. We show that new drum classes can be learned and performance for known classes can be improved by providing some examples of that respective class during test time.enmusic information retrievalautomatic drum transcriptionfew-shot learningreal-time processingconference paper