Dynamic Percussion Dataset

This dataset was recorded as a part of my Bachelor's thesis work for Tampere University. It is a collection of samples from different drums and percussion, recorded in an anechoic room. More details on the structure, labels and recording process can be found in [1]. In the thesis, this dataset was visualized in the 2-dimensional space with a method (described in [1]) that utilizes OpenL3 deep audio embeddings [2] and t-SNE dimensionality reduction algorithm [3]. Since the embeddings can take a significant time to compute without a sufficient CUDA environment installed, they are provided with the dataset (in a later version). <strong>Audio format:</strong><br> Data type: wav<br> Sampling rate: 44,1 kHz<br> Bit-depth: 24<br> Channels: 1 <em>Added in version 1.1.0:</em> OpenL3 pre-calculated embeddings have been added. Several different models were used to calculate the embeddings. Each file corresponds to a single pickle file containing the embeddings. The pickle files contain a metadata dictionary which has information about the original sample and the OpenL3 model used to calculate that particular embedding series. NOTE: The reference [1] will be added when the thesis has been published.