Deploying deep learning models on embedded devices is an arduous task: oftentimes, there exist no platform-specific instructions, and compilation times can be considerably large due to the limited computational resources available on-device. Moreover, many music-making applications demand real-time inference. Embedded hardware platforms for audio, such as Bela, offer an entry point for beginners into physical audio computing; however, the need for cross-compilation environments and low-level software development tools for deploying embedded deep learning models imposes high entry barriers on non-expert users. We present a pipeline for deploying neural networks in the Bela embedded hardware platform. In our pipeline, we include a tool to record a multichannel dataset of sensor signals. Additionally, we provide a dockerised cross-compilation environment for faster compilation. With this pipeline, we aim to provide a template for programmers and makers to prototype and experiment with neural networks for real-time embedded musical applications.

翻译：在嵌入式设备上部署深度学习模型是一项艰巨的任务：通常没有特定平台的指令，且由于设备上计算资源有限，编译时间可能相当长。此外，许多音乐制作应用需要实时推理。用于音频的嵌入式硬件平台（如Bela）为初学者提供了进入物理音频计算的入门点；然而，部署嵌入式深度学习模型所需的交叉编译环境和底层软件开发工具，对非专业用户构成了高门槛。我们提出了一套在Bela嵌入式硬件平台上部署神经网络的流程。在该流程中，我们包含了一个工具，用于记录多通道传感器信号数据集。此外，我们提供了一个Docker化的交叉编译环境，以加快编译速度。通过这一流程，我们旨在为程序员和创客提供一个模板，用于原型设计和实验面向实时嵌入式音乐应用的神经网络。