Machine learning based noise suppression in narrow-band speech communication systems

Narrow-band digital personal radio systems are used for speech communication in challenging environments where background noise, such as machinery or emergency sirens, can pose significant problems for speech intelligibility. This paper proposes a machine learning based noise suppression approach that utilises a neuro-fuzzy logic-based neural network for noise estimation and reduction. The technique is shown to give significant improvements in noise suppression compared to a non-adaptive noise suppression approach. The choice of a neuro-fuzzy logic neural network is motivated by the need for a low-power implementation suitable for mobile, power constrained, terminals. To validate this, the algorithm has been tested in a real-time system showing that it can be implemented in constrained devices unlike more complex machine learning techniques that are unsuitable for low-power digital personal radio systems.