摘要
环境声音分类(Environmental Sound Classification,ESC)是非语音音频分类任务最重要的课题之一。近年来,深度神经网络(Deep Neural Network,DNN)方法在ESC方面取得了许多进展。然而,DNN是计算和存储密集型的,无法直接部署到基于微控制器(Microcontroller Unit,MCU)的物联网设备上。针对这一问题,本文提出一种用于资源高度受限设备的DNN压缩方法。由于DNN模型参数规模较大无法直接部署,因此提出使用剪枝方法进行大幅压缩,并针对该操作带来的精度损失问题,设计一种基于模型中间层特征信息的知识蒸馏方法。基于STM32F746ZG设备在公开的数据集(UrbanSound8K、ESC-50)上进行测试,实验结果表明,本文方法能够获得高达97%的压缩率,同时保持良好的推理精度和速度。
Environmental Sound Classification(ESC)is known as one of the most important topics of the non-speech audio clas sification task.In recent years,deep neural networks(DNNs)have made a lot of progress in ESC.However,DNNs are computa tionally and memory-intensive,and cannot be directly deployed on IoT devices based on microcontroller units(MCU).To ad dress this problem,this paper proposes a DNN compression method for highly resource-constrained devices.Since DNNs have a large number of parameters,which cannot be directly deployed,so this paper proposes to use the pruning method for substantial compression.Afterwards,aiming at the problem of accuracy loss caused by this operation,we design a knowledge distillation based on the feature information of multiple intermediate layers.Tests are carried out on public datasets(UrbanSound8K,ESC-50)using the STM32F746ZG device.The experimental results demonstrate that proposed method can achieve up to 97%com pression rate while maintaining good inference performance and speed.
作者
孟娜
方维维
路红英
MENG Na;FANG Wei-wei;LU Hong-ying(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China)
出处
《计算机与现代化》
2024年第1期80-86,共7页
Computer and Modernization
关键词
环境声音分类
边缘计算
微控制器
剪枝
知识蒸馏
量化
environmental sound classification
edge computing
microcontroller unit
pruning
knowledge distillation
quanti zation
