Page 39 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)
P. 39

Dinh-Hoang-Long Tran, Quoc-Huy Le                                                23

                        Implementation of Convolutional Neural Network on

                         a Microcontroller for Classification of Audio Signals


                                                                                 *
                                           Dinh-Hoang-Long Tran, Quoc-Huy Le
                              Faculty of Advanced Science and Technology, The University of Danang -
                                  University of Science and Technology, Danang 550000, Vietnam.
                                                   *  lqhuy@dut.udn.vn




                           Abstract. The goal of this work is to develop a compact and low-cost device to
                           detect dangerous and suspicious sounds in a sensitive area. The proposed solution
                           uses  an  STM32  microcontroller  embedded  with  a  deep  learning  model  and
                           equipped  with  various  peripherals.  For  the  demo  purpose  we  used  the
                           STM32F746NGH6  Discovery  Kit  and built  the  convolutional  neural  network
                           embedded in this microcontroller with the popular Keras API. We illustrated that
                           the deep learning model using convolutional neural networks algorithm can be
                           implemented  on  STM32F746NGH6  microcontroller  kit  and  the  device  can
                           classify the labeled audio with an accuracy of about 97%. We also found that
                           signals from real sound sensors or real microphones have noises which affects
                           strongly on the model accuracy and thus it is necessary to build the dataset based
                           on  the  available  hardware.  Our  in-progress  work  is  to  record  audio  using
                           ADMP401  analog  MEMS  microphone  available  on  the  STM32F746NGH6
                           microcontroller and then using these datasets for a second experimental model.

                           Keywords:  Audio  Classification,  Deep  Learning,  Convolutional  Neural
                           Networks, Edge AI, Audio Signal Processing.



                     1     Introduction


                     Sound  signals  are  valuable  source  of  information [1][2].  Physically,  sound  is  a
                     mechanical movement in a continuous medium, each vibration produces energy with
                     its characteristics based on characteristics such as frequency, and amplitude to identify
                     the source emitting it. Every action, every movement, and the phenomenon around our
                     environment creates vibrations from simple actions such as walking and talking. We
                     can build a system that makes good use of sound to detect dangerous events, e.g. when
                     a robber threatens a staff or infiltrator tries break the window. Using deep learning we
                     can have a model that could classify incoming sound to detect that sound is safe or
                     suspicious,  thereby  detecting  a  dangerous  event  [3][4][5][6].  This  model  can  be
                     then  embedded  to  a  microcontroller  to  build  a  system  for  automatic  detection  and
                     classification of sounds [7][8][9]. The idea of this work is to build such a system that
                     can be used in warning system as illustrated in Fig. 1.









                     ISBN: 978-604-80-8083-9                                                  CITA 2023
   34   35   36   37   38   39   40   41   42   43   44