Efficient End-to-End Audio Embeddings Generation for Audio Classification on Target Applications

We describe a general-purpose end-to-end audio embeddings generator that can be easily adapted to various acoustic scene and event classification applications. In contrast to many other models for audio classification, this end-to-end embeddings generator does not require a separate feature extraction step, but processes audio samples directly which simplifies its porting into hardware platforms. Our approach learns a generic audio embedding representation that is pre-trained on a large audio dataset. It can then be fine-tuned via transfer learning with limited data requirements for an audio classification target application. We compare different transfer learning strategies and evaluate the proposed embeddings generator on three popular publicly available datasets, and report results for different model sizes. The experimental results show that this end-to-end approach presents a scalable, efficient, and competitive alternative to more common spectral-based audio classifiers.