Vibrary: A Consumer-Trainable Music Tagging Utility

摘要

We present the engineering underlying a consumer application to help music industry professionals find audio clips and samples of personal interest within their large audio libraries typically consisting of heterogeneously-labeled clips supplied by various vendors. We enable users to train an indexing system using their own custom tags (e.g., instruments, genres, moods), by means of convolutional neural networks operating on spectrograms. Since the intended users are not data scientists and may not possess the required computational resources (i.e. Graphics Processing Units, GPUs), our primary contributions consist of i) designing an intuitive user experience for a local client application to help users create representative spectrogram datasets, and ii) 'seamless' integration with a cloud-based GPU server for efficient neural network training.