Page 47 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

Dinh-Hoang-Long Tran, Quoc-Huy Le 31

In Summary, environment class have the highest recall and precious, it may be due to
the dataset, because this sound was recorded by the same devices. While other labels
were taken from the internet which mean it was record by various device, different
quality led to lower statistics. However, these results still prove a well performance of
the model since all 3 class has recall and precious higher than 0.9 except the glass label.
Testing in real life: the proposed model was embedded to STM32 and was tested in
real condition to evaluate its performance. We use the AMDP401 for collecting input
signal and a speaker to play various type of sound. The result was transmitted from
STM32F746 Kit to computer screen via UART communication protocol (Fig. 9). The

between gunshot and glass breaking sound.

Fig. 9. Transmit the result via UART

4 Conclusion

In this work we demonstrated that a deep learning model using convolutional neural
net-works algorithm can be implemented on the low-cost STM32F746NGH6

ISBN: 978-604-80-8083-9 CITA 2023