Low-Power AI Solution for Intelligent Audio Scene Classification
Audio Scene Classification (ASC) is a technology that enhances device intelligence by enabling environmental awareness. It has wide-ranging applications in wearable devices, security, environmental monitoring, and healthcare. However, implementing ASC presents challenges in software development and hardware design, particularly for portable or wearable devices with processing, memory, and power constraints. To address these issues, STMicroelectronics offers a cost-effective, low-power AI solution based on STM32 microcontrollers, combining efficient processing with minimal power consumption.
System Architecture and Key Components
This solution is built on the B-L475E-IOT01A Discovery Kit, which runs the FP-AI-SENSING1 function pack to execute edge-based ASC tasks on the STM32L4 microcontroller.
Two MP34DT05 digital omnidirectional MEMS microphones capture environmental audio, ensuring high-quality sound detection. The STM32L475VG MCU, known for its ultra-low power consumption, features signal processing peripherals and a Floating Point Unit (FPU) for efficient AI execution.
AI Processing Workflow
ASC tasks leverage Artificial Neural Networks (ANN) for inference, following this process:
Audio Data Acquisition – MEMS microphones capture ambient sounds and send the data to the STM32L4 MCU.
Data Preprocessing – Audio samples are buffered and enter the ASC preprocessing stage, which includes Fast Fourier Transform (FFT), filter bank application, and logarithmic scaling to extract spectral features.
Neural Network Inference – The extracted audio features are fed into an ASC Convolutional Neural Network (CNN), which classifies the scene as indoor, outdoor, or in-vehicle at a rate of one inference per second.
Minimal Storage Requirement – The ANN requires only 18KB RAM and 31KB Flash, eliminating the need for external memory and making it ideal for embedded systems.
Wireless Connectivity & Smartphone Interaction
For wireless data transmission, the system integrates the STM32WB55VGY MCU, which supports Bluetooth Low Energy (BLE) 5.2. The ASC results can be transmitted to a smartphone running the ST BLE Sensor app (Android & iOS), allowing users to visualize classification outputs and activate data logging for AI retraining and optimization.
Key Advantages & Application Prospects
This AI solution offers several key benefits:
Low Power & High Efficiency – STM32L4 and STM32WB55VGY ensure ultra-low power operation while delivering high-performance AI processing.
Minimal Memory Footprint – With only 18KB RAM and 31KB Flash, the system eliminates the need for external storage, making it ideal for embedded applications.
Wireless Data Transmission – BLE 5.2 connectivity enables seamless smartphone integration for real-time data monitoring and logging.
Versatile Applications – This technology can be widely applied in smart homes, wearable devices, environmental monitoring, and security systems, enhancing devices with real-time environmental awareness.
Conclusion
This cost-effective, low-power AI solution based on STM32 enables real-time audio scene classification using neural network technology. With BLE connectivity, it allows seamless interaction with smartphones, providing intelligent environmental awareness for embedded devices. This solution paves the way for AI-driven edge computing applications, enhancing the capabilities of smart devices across multiple industries.
For inquiries or to explore how these solutions can benefit your projects, feel free to contact ICHOME.
