Neuro-Fuzzy voice quality improvement for low-power, half-duplex, communication

Half-duplex mobile radio systems are commonly used for speech communication, often in challenging environments where background noise may prevent speech intelligibility. These systems usually operate with modest battery capacity which strictly limits the available computational power. In this paper, we present a noise suppression approach that uses lightweight machine learning. The machine learning leverages a neuro-fuzzy logic-based neural network to create accurate noise estimation that is used adaptively to create a filter for noise reduction. The system buffers a number of noise samples, triggered by the half-duplex key press, allowing the solution to adapt to recent changes in the background auditory noise. The selection of a neuro-fuzzy logic-based neural network is driven by the necessity for a low-power implementation suitable for mobile, power-constrained terminals. The proposed system is compared with both traditional noise suppression methods as well as with a deep convolutional neural network (DCNN) variant of our system. The results demonstrate that our system significantly outperforms traditional methods in terms of both subjective and objective quality metrics. Further, they show that, although the DCNN variant achieves comparable performance, it has a high computational cost which renders it unsuitable for low-power digital personal radio systems. To validate its practicality, the proposed method is tested in a real-time system, demonstrating its compatibility with constrained devices.