基于GD32的低功耗语音唤醒模块设计与实现

Design and Implementation of Low-power Voice Wake-up Module Based on GD32

摘要:

唤醒模块可以为机器人等长期待机运行的设备提供快速唤醒、降低功耗等功能；低功耗及智能化是评价该类模块的重要性能指标；设计了一款基于GD32的低功耗语音唤醒模块,该模块采用微型机器学习(Tiny Machine Learning, TinyML)技术进行语音识别,将部分机器学习运算转移到单片机上进行；在电脑端对音频文件进行预处理,并将处理后的音频通过短时傅里叶变换转化为频谱图,进行模型训练；将训练好的模型由TensorFlow Lite转化为单片机可使用的C语言数组,通过微控制器开发环境将数组部署到GD32芯片上；以婴儿哭声检测为例,经实验测试,模块识别准确率约在70%左右,模型和数据集仍有进一步优化的空间；对模块的功耗也进行了测试与优化；该模块的设计与实现对微控制器实现机器学习具有一定的参考意义。

Abstract:

Wake-up module can provide fast wake-up, reduce power consumption and other functions for robots and other devices running on standby for a long time; low power consumption and intelligence is an important performance index for evaluating this kind of module; a low-power voice wake-up module based on GD32 is designed, which adopts the Tiny Machine Learning technology for speech recognition, and transfers part of the Machine Learning operations are transferred to the microcontroller; pre-process the audio file on the computer side, convert the audio into a spectrogram by short-time Fourier transform, and train the model; the trained model is converted by TensorFlow Lite into a C language array that can be used by the microcontroller, and deployed on the GD32 chip; take the baby cry detection as an example, after experimental testing, the module recognition accuracy rate is About 70%, the model and dataset still have room for further optimization; the power consumption of the module is tested and optimized; the design and implementation of the module has a certain reference significance for the implementation of machine learning in microcontrollers.