INFDok - Dokumentenvolltextserver des Fachbereichs Informatik

Kurzfassung auf Englisch:

In this work, we investigate the applicability of the Online Deep Attractor Network (ODANet) for real-time speech separation on an embedded system with limited processing resources. To optimize the ODANet for a resource constrained environment, we extensively evaluate two different reduction methods. First, we present a detailed analysis of complexity reduction via hyper-parameter tuning of the ODANet and second, we introduce a compression method for long short-term memory (LSTM) layers to the ODANet architecture. While our results suggest that real-time capability is possible for a desktop computer with these methods, it is not achievable for an embedded device like the NVIDIA Jetson Nano while maintaining an acceptable separation performance. In further findings, we show that the utilized compression method for LSTMs is superior to hyper-parameter tuning in terms of finding a good trade-off between low processing time and separation performance. Furthermore, we want to highlight that this work is the first to our knowledge to give an extensive description of a singular value decomposition based compression method for LSTMs including an open-source implementation available at https://github.com/sp-uhh/compressed-lstm .