Real-Time Automotive Engine Sound Simulation with Deep Neural Network

This paper introduces a real-time technique for simulating automotive engine sounds based on revolutions per minute (RPM) and pedal pressure data. We present a hybrid approach combining both sample-based and procedural methods. In the sample-based technique, the sound of an idle engine undergoes pitch-shifting proportional to the ratio of current RPM to idle RPM. For the procedural technique, deep neural networks fine-tune the amplitude of the engine’s pulse frequency derived from the sample-based sound. To ensure the synthesized sound does not have any clicks between the frames, we utilize a modified griffin-lim algorithm at the frame level, which, with our proposed overlap-and-add feature, can bridge the phase gap between two frames. Experimental evaluations on our self-collected database validate the efficacy of the introduced approach.