Throughout its history, the engine acoustic character has been emblematic of the product essence, owing to its robust correlation of factors like in-cylinder pressure gradients, components design, and perceived quality. Best practice for engine acoustic characterization requires the employment of a hemi-anechoic chamber, a significant number of sensors and special acoustic insulation for engine ancillaries and transmission. This process is highly demanding in terms of cost and time due to multiple engine working points to be tested and consequent data post-processing. Given the apparent underutilization of Neural Networks (NN) predictive capabilities in this research area, the following paper introduces a tool capable of estimation of engine acoustic performance by processing system inputs (e.g., Injected Fuel, Rail Pressure). This is achieved through the application of a Multi-Layer Perceptron (MLP), which operates as a feedforward network functioning at stationary points. In particular, the investigation addresses the estimation of direct Combustion Noise (CN), Sound Power (PWL) averaged over the main radiating surfaces, Loudness and Modulation. The Neural Network was trained and tested under low and medium load/speed operating conditions of an inline 4-cylinder turbocharged diesel engine. The models achieve less than 0.5% of Test Root Mean Square Error (RMSE) for the estimation of CN and Sound Power, less than 2% Test RMSE for Loudness and less than 4% for Modulation. In addition, the same training procedure and network architectures were used to predict the third of octaves quantities (CN, Sound Power and Loudness) with a slightly decrease of model accuracy.