Real-Time Generation of a Targeted Clean Audio Sequence from Source Separation of Noisy Environmental Mixtures Using a Deep Nonnegative Matrix Factorization on IOT Devices
摘要
Real-time acquisition and processing of audio signals is a growing field in ICT, mainly for monitoring applications. The availability of miniaturized microphones and microcontrollers with hardware floating-point units, enables new methods for real-time audio processing that were impossible a few years ago in IOT devices. In this paper we present an algorithm for real-time extraction and reproduction of targeted clean sources from noisy environmental mixtures. Thanks to a scientific machine learning approach we obtain results well beyond classical filtering.