Investigation of Machine Learning Methods for Colour Audio Noise Suppression

Separating background noise or suppressing it from a given recording to extract clean speech benefits humans with hearing aids and machines, such as Automatic Speech Recognition systems. In recent years, various experiments have been conducted to develop a Speech Enhancement (SE) model or framework that would produce high-quality estimations based on classical algorithms or Deep Neural Network methods. Every research tries to evaluate and compare the new approach with existing ones. Still, a comparison is often biased or insufficient because of the different setup of training data or evaluation metrics used between each trial. The criteria used to corrupt the data with additive noise differ per research, like the evaluation strategy. This work investigates the most promising SE approaches by replicating experiments and using exact training and testing criteria and the same set of evaluation metrics to identify what produces better clean speech signal estimations.