Reference / Literaturverweis [Bre22b]

Hagen, F., Hartig, J., Helmich, R., Mahnke, M., Schade, D.-N., Sommerfeld, F., Stöver, H., Wittje, P., Bredereke, J.:
Neural Network on an FPGA for Speech Command Recognition on an Autonomous Vehicle.
Technical Report. City University of Applied Sciences Bremen, Germany (Mar. 2022).

Abstract / Zusammenfassung

This research focuses on solutions for executing Neural Networks (NNs) on field programmable gate arrays (FPGAs) of comparably limited power as efficiently as possible, for use of such networks on autonomous vehicles.

Specifically, we propose a concept of a speech command recognition using a system-on-chip (SoC), where we implement the first data processing stages on the FPGA of the SoC. A proof of concept for acquiring audio data via the FPGA and converting it to Mel Frequency Cepstral Coefficients (MFCCs) has been made. Using a Quantized Neural Network (QNN) we have reached an accuracy of 90% on an audio test set for driving commands. However the NN yields worse results during inference. The general concept of the data processsing pipeline in this research, enables researchers to proceed the implementation on FPGA hardware. As demonstrated by the developed NN speech recognition is possible on comparative performance-limited FPGA boards.

Full Text / Volltext

PDF (4.7 MB)

Source Code / Quelltexte

Please just ask me for a ZIP file with the source code, if you need it. / Bei Bedarf fragen Sie mich bitte einfach nach einer ZIP-Datei mit den Quelltexten.