a) Separation of Motorcycle Sound by Near Field Microphone Array and Nonnegative Matrix Factorization Chisaki YOSHINAGA, Nonmember, Yosuke TATEKURA a), Member, Kazuaki HAMADA, and Tetsuya KIMURA, Nonmembers Shizuoka University, Hamamatsu-shi, 432 8561 Japan Yamaha Motor Co., Ltd., Iwata-shi, 438 8501 Japan a) E-mail: tatekura.yosuke@shizuoka.ac.jp 1. / / (Delay and Sum: DS) [1] (Spectral Subtraction: SS) [2] [3] [3] (Nonnegative Matrix Factorization: NMF) [4] 2. 0.5 m 0.5 m 1: fig1.wav 1 1 A B 2 A B 1 Fig. 1 An example of time waveform of the motorcycle sound. 232 A Vol. J98 A No. 2 pp. 232 236 c 2015
2 DS SS (SS+DS) Fig. 2 Separation algorithm for motorcycle sound by combining DS and SS (SS+DS). 3. [3] DS SS 2 SS+DS 1 1 B SS SS 4. NMF 4. 1 (NMF) Y =[y fn ] F N Y HU (1) H =[h fk ] F K U = [v kn ] K N f (= 1,...,F) n (= 1,...,N) k (= 1,...,K) [a ij] I J i j a I J 3 SS+DS NMF Fig. 3 Separation algorithm by combining SS+DS and NMF. H U (1) H U NMF H U 4. 2 NMF 3. SS+DS NMF 3 3. ŝ 1(t) ŝ 2(t) / ŝ 1(t), ŝ 2(t) NMF H ŝ 1(t), ŝ 2(t) NMF 3. ŝ i(t) ŝ i(t) =[ŝ i(1) ŝ i(2) ŝ i(n)] T (i =1, 2) Ŝi(f) =[Ŝi(1) Ŝi(2) Ŝ i(n)] T (i =1, 2) / K 1,K 2 (K = K 1 + K 2) Ĥ 1 [ ] Ĥ 1 = Ŝ 1(f) Ŝ 1(f) (2) } {{ } F K 1 Ŝi(f) K1 233
2015/2 Vol. J98 A No. 2 Ĥ 2 NMF H (0) [ ] H (0) = Ĥ 1 Ĥ 2 (3) } {{ } F K / / 5. 5. 1 4 8 0.50 m 0.50 m 0.23 m Mic. 1,, Mic. 8 SS+DS 8 NMF Mic. 6 Mic. 6 48000 Hz 1024 512 Hanning NMF 100 2 K 1 K 2 50 (1) Kullback-Leibler divergence 5. 2 5. 2. 1 SS+DS SS+DS NMF / SS+DS NMF NMF SNR 5 6 NMF 1 B SS+DS NMF 4 Fig. 4 Arrangement of the microphones for recording motorcycle sound, in which the dashed lines indicate height from the ground. Fig. 5 5 NMF Time waveform of separated exhaust sound before/after NMF. 234
Fig. 6 6 NMF Time waveform of separated mechanical sound before/after NMF. NMF 5. 2. 2 2: fig5a.wavfig5b.wav NMF 3: fig6a.wavfig6b.wav NMF NMF NMF NMF NMF / NMF NMF / / 5. 2. 3 /NMF 20 6 1 fig1.wav 5 7 Fig. 7 NMF Result of the subjective evaluation for the exhaust sound before/after NMF. 5: 4: 3: 2: 1: NMF 2 fig5a.wav fig5b.wav 2 7 Mean opinion scoremos NMF MOS 5% t t(11) = 3.46P <0.01 5 5: 4: 3: 2: 1: NMF 3 fig6a.wav fig6b.wav 2 8 MOS NMF MOS 5% 235
2015/2 Vol. J98 A No. 2 8 Fig. 8 Fig. 9 NMF Result of the subjective evaluation for the mechanical sound before/after NMF. 9 SNR Comparison of reconstructed sound by SNR. t t(11) = 9.10P <0.01 5. 2. 4 SNR / SNR t SNR [db] = y2 (t) (4) t [y2 (t) {e N(t)+m N(t)}] 2 y(t) Mic.6 e N(t) NMF m N(t) NMF SNR 9 NMF SNR 20dB SS+DS NMF / 6. DS+SS NMF / / [1] J. Benesty, W. Kellermann, Eds., Microphone Array Signal Processing, Springer, Berlin, 2008. [2] S.F. Boll, Suppression of acoustic noise in speech using spectral subtraction, IEEE Trans. Acoust. Speech Signal Process., vol.assp-27, no.2, pp.113 120, 1979. [3] 57 SCI 13, 114-2, 2013. [4] NMF / vol.95, no.9, pp.829 833, Sept. 2012. A 1 1 Table A 1 Attached data 1. fig1.wav A 2 2 Table A 2 Attached data 2. fig5a.wav, fig5b.wav NMF A 3 3 Table A 3 Attached data 3. fig5a.wav, fig5b.wav NMF 26 4 3 8 13 236