Improved Monaural Speech Enhancement via Low-Complexity Fully Connected Neural Networks: A Performance Analysis

In recent times, both conventional and Deep Neural Network (DNN)-based approaches have made successful advancements in speech enhancement. We consider one of the recent conventional approaches, the image analysis technique (IAT), and some DNN-based approaches, the convolutional neural network (CNN), and the genetic algorithm-DNN (GA-DNN) employed for speech enhancement. These architectures including the CNN have a higher run-time computational complexity relative to a simple fully connected neural network (FCNN) given we have used floating-point operations (FLOP) as the metric. In this paper, we attempt to demonstrate how an FCNN shows results on par with or superior to some of the compared algorithms despite having lower run-time computational complexity. The complexity comparisons have been shown using FLOP count as the metric along with multiple measures to ensure speech quality and intelligibility. Short-term objective intelligibility and perceptual evaluation of speech quality are the objective measures used for comparison. Additionally, the modified rhyming test and the mean opinion score are used as subjective intelligibility measures. Comparing the average run times of the algorithms, we get a run time of 0.1 s per speech signal for the low complexity DNN on a standard laptop computer which is far superior to the 65 s taken per speech signal in the non-DNN approach making it more suitable for real time deployment. The experimental results also show comparisons between the implemented algorithm and its counterparts in the cases of stationary noise, non-stationary noise, and interfering speech.