Classifying environmental sounds using image recognition networks

类型	说明
论文信息	Classifying environmental sounds using image recognition networks Venkatesh Boddapatia, Andrej Petefb, Jim Rasmussonb, Lars Lundberg
会议期刊	International Conference on Knowledge Based and Intelligent Information and Engineering Systems, KES2017, 6-8 September 2017, Marseille, France

介绍	说明
简介	1.环境声音自动分类有助于智能家居、远程监控的应用 2.现有移动设备普遍有图像深度神经网络模型，本文研究用现有图像分类模型Alexnet、GoogLeNet模型进行语音分类

常用音频预处理方法	说明
(1) Framing-based	1) Framing-based where the audio signals are separated into frames using a Hamming window. Then the features are extracted from each frame and classified separately.
(2) Sub-framing	2) Sub-framing based processing where the frames are further subdivided and each frame is classified based on the majority voting of the sub-frames.
(3) Sequential processing	3) Sequential processing where the audio signals are divided into segments of typically 30 ms with 50% overlap. The classifier then classifies the features extracted from these segments.

数据集	说明
(1)ESC-10	400个5秒音频，10类，每类40个音频
(2)ESC-50	2000个5秒音频，50类，每类40个音频
(3)UrbanSound8K	8732个<=4s的环境声音音频，共10类， {0: 1000, 1: 429, 2: 1000, 3: 1000, 4: 1000, 5: 1000, 6: 374, 7: 1000, 8: 929, 9: 1000}

方法	说明
(1)特征	MFCC(Mel-Frequency Cepstral Coefficients)、Spectrogram、CRP(Cross Recurrence Plot)三类特征单通道图以及组合成的三通道图(256×256) )
(2)模型	AlexNet、GoogLeNet
(3)设置	五折交叉验证、32kHz采样率、帧长30ms、帧重叠率50%、50训练周期、学习率0.01、SGD优化器、学习率计划：指数(0.95)衰减、256×256

实验	说明
(1)三类特征两种模型实验	****
(2)采样率实验	最优采样率分别用于后续实验
(3)帧长实验	)
(4)UrbanSound8K实验以及CRNN实验	CRNN并没有提高准确度

结论	说明
①	在ESC-50、ESC-10和UrbanSound8K数据集上，GoogLeNet的分类准确率分别为73%、91%和93%。
②	三通道组合图并没有提高分类准确率、CRNN也没有得到较高精度

---

1. 1.paper ↩