Convolutional Neural Network Based Indoor Microphone Array Sound Source Localization

Chen Jiao; Tao Zhang; Jianhong Sun

doi:10.3788/LOP57.081021

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 8 >Page 081021 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 8, 081021 (2020)

Convolutional Neural Network Based Indoor Microphone Array Sound Source Localization

Chen Jiao^*, Tao Zhang, and Jianhong Sun

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/LOP57.081021 Cite this Article Set citation alerts

Chen Jiao, Tao Zhang, Jianhong Sun. Convolutional Neural Network Based Indoor Microphone Array Sound Source Localization[J]. Laser & Optoelectronics Progress, 2020, 57(8): 081021 Copy Citation Text

show less

Fig. 1. Space cluster classification for SSL

Download full size

Fig. 2. Flow chart of localization method

Download full size

Fig. 3. Comparison of classification accuracy of different algorithms in different environments

Download full size

Network layer	Network parameter
Input layer	Dimension: 112×2688
Convolution layer	Number of convolution kernel: 8; kernel_size: 5×5; stride: 1; pad: 0
Pooling layer	Pooling: max pooling; kernel_size: 3×3; stride: 1; pad: 0; dropout: 50%
Convolution layer	Number of convolution kernel: 16; kernel_size: 5×5; stride: 1; pad 0
Pooling layer	Pooling: max pooling; kernel_size: 3×3; stride: 1; pad 0; dropout: 50%
Convolution layer	Number of convolution kernel: 32; kernel_size: 5×5; stride: 1; pad 0
Pooling layer	Pooling: max pooling; kernel_size: 3×3; stride: 1; pad 0; dropout: 50%
Convolution layer	Number of convolution kernel: 64; kernel_size: 5×5; stride: 1; pad 0
Pooling layer	Pooling: max pooling; kernel_size: 3×3; stride: 1; pad 0; dropout: 50%
Connection layer	Number of neurons: 1024; activation function: ReLU; dropout: 50%
Output layer	Activation function: softmax; learning rate : 0.001; iterations: 1000; batch_size: 64

Table 1. CNN structure

Signal	Reverberationtime /ms	Accuracyof TDOA /%	Accuracyof SVM /%	Accuracyof PNN /%	Accuracyof BP /%	Accuracyof CNN /%
Clean voice	0	62.61	78.58	96.87	96.32	96.67
Clean voice	300	61.64	75.28	96.30	95.53	95.03
Clean voice	600	60.02	74.86	93.25	92.84	93.80
SNR: 10 dB	0	42.92	44.93	90.16	94.08	95.83
SNR: 10 dB	300	46.31	49.82	89.44	90.36	93.87
SNR: 10 dB	600	36.53	46.35	88.73	87.69	92.79
SNR: 0 dB	0	38.61	45.57	89.89	88.81	94.32
SNR: 0 dB	300	34.61	44.03	89.09	86.37	93.49
SNR: 0 dB	300	24.87	44.60	88.61	85.20	90.78

Table 2. Classification accuracy of different algorithms in different environments

Signal	Reverberationtime /ms	Iterations	Localizationtime /s	Iterations	Localizationtime ofPNN /s	Iterations	Localizationtime ofBP /s	Iterations	Localizationtime ofCNN /s
Clean voice	0	4496	7.4	4671	6.8	10504	10.3	18723	11.2
Clean voice	300	4215	7.8	5387	7.1	13661	9.8	19025	10.8
Clean voice	600	4615	8.1	6062	6.1	22572	10.3	29781	10.9
SNR: 10 dB	0	4853	9.3	5283	7.4	21973	9.7	33671	12.4
SNR: 10 dB	300	49.82	8.2	6231	7.3	34603	9.7	42101	11.8
SNR: 10 dB	600	5430	9.1	5735	7.1	38524	10.9	48776	11.2
SNR: 0 dB	0	4551	9.2	5089	6.9	35871	10.6	54786	11.2
SNR: 0 dB	300	4876	9.4	6086	8.3	41654	10.8	57623	10.5
SNR: 0 dB	300	5576	9.3	6487	8.1	47726	10.8	58341	11.9

Table 3. Real-time localization time of different algorithms

Chen Jiao, Tao Zhang, Jianhong Sun. Convolutional Neural Network Based Indoor Microphone Array Sound Source Localization[J]. Laser & Optoelectronics Progress, 2020, 57(8): 081021

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information