8-439 - Deep Embeddings with Essentia Models

Pablo Alonso-Jiménez, Dmitry Bogdanov, Xavier Serra

Abstract: We present the integration of various CNN TensorFlow models developed for different MIR tasks into Essentia. This is a continuation of our previous work, extending the list of supported models and adding new algorithms to facilitate usability. Essentia provides input feature extraction and inference with TensorFlow models in a single C++ pipeline with Python bindings, the overhead of Python bindings and facilitating the deployment of C++ and Python MIR applications. We assess the new models' capabilities to serve as embedding extractors in many downstream classification tasks. All presented models are publicly available on the Essentia website.