2_05.dvi

Size: px

Start display at page:

Download "2_05.dvi"

えいじろうたみや
5 years ago
Views:

1 pp c * 1, 2 1 2, BM CSS CSS CSM BM CSM CSS CSS CSM Blind source separation, Sparseness, Binary mas, Musical noise, Cepstral smoothing, Separated speech signals 1. BSS [1] BSS BSS ICA [2] e.g, [3] N M BSS Cepstral smoothing of separated signals for underdetermined speech separation, by Yumi Ansai, Shoo Arai, Shoji Maino, Tomohiro Naatani, Taeshi Yamada, Atsushi Naamura and Nobuhio Kitawai. 1 NTT BM [3 15] BM BSS BM [16] BM CSM: Cepstral Smoothing of spectral Mass CSM CSM BM CSM

2 75 [16] BM BM CSS: Cepstral Smoothing of separated Signals CSM [17] CSM CSS CSS 2 BM BSS 3 BM CSM 4 CSS 5 CSS CSM s i (i =1,...,N) M j x j (j =1,...,M) N L x j (n) = h ji (l)s i (n l +1) i=1 l=1 (j =1,...,M) (1) h ji i j l n 1 BSS x j y i N >M N M [7, 18 21] 1 BSS N >M N X j (f,m) = H ji (f)s i (f,m) i=1 (j =1,...,M) (2) H ji (f) i j S i (f,m) X j (f,m) STFT f m 2.2 BM [3, 6] 2 N -means X(f,m) =[X 1 (f,m),...,x M (f,m)] T [6] { 1 X(f,m) C, M (f,m) = (3) otherwise M min C M min 0 (> 0) BM Y (f,m) =M (f,m)x j (f,m) (4) Y (f,m) STFT y BM [15] CSM BM CSM

3 l pitch =argmax{m cepst (l, m) l low l l high } l (8) {l low,l high } Hz [16] (8) 2 BM CSM [16] 2 CSM M cepst (l, m) = DFT 1 {ln(m (f,m)) f=0,...,f 1 } (5) l DFT{ } F M (f,m) (3) M min =0.01 M cepst M cepst (l, m) =β l M cepst (l, m 1) +(1 β l )M cepst (l, m) (6) (6) l β l β env if l {0,...,l env } β l = β pitch if l = l pitch β pea if l {(l env +1),...,F/2}\{l pitch } (7) F (5) F [16] l {0,...,l env } M cepst (l, m) M (f,m) β env M (f,m) l = l pitch β pitch M (f,m) β pea (>β pitch ) m l pitch l pitch l pitch (7) l>f/2 DFT M cepst (l, m) DFT M (f,m) M (f,m) =exp(dft{m cepst (l, m) l=0,...,f 1 }) (9) Y (f,m) =M (f,m)x j (f,m) (10) 3.2 CSM 3 BM CSM M =3 N =4 BM CSM 4 BM 3(A) CSM 3(B) CSM CSM BM BM 1 M min 2 4 (A) BM (B) (C)(D) M cepst 4(C) 4(A) BM 4(B) (7)

4 劣決定音源分離のための分離音声のケプストラムスムージング図 3 分離信号のスペクトログラム 77

5 CSM (A) (B) BM (C) CSM (D) CSM CSS 3.2 CSM CSS

6 79 1 [16] f s =8Hz l env =16 β env =0 F = 512 l low =32 β pitch =0.4 M min =0.01 l high = 228 β pea =0.8 5 CSS 4.2 β pea 5 CSS 2 CSS CSM 4.2 CSS (4) Y (f,m) Y cepst (l, m)=dft 1 {ln(y (f,m)) f=0,...,f 1 } (11) Y cepst Y cepst (l, m) =β l Y cepst (l, m 1) +(1 β l )Y cepst (l, m) (12) β l (7) l {(l env + 1),...,F/2}\{l pitch } (12) β pea l pitch l pitch =argmax{y cepst (l, m) l low l l high } l (13) 2 β l original case 1 case 2 case 3 CSM β pitch β pea CSS β pitch β pea [16] l pitch l > F/2 DFT Y cepst (l, m) Y cepst (l, m) (9) DFT STFT y CSS CSM 1 CSS CSM CSM β l [16] 1 CSS β l 3 β l 2 β env 0 β pitch β pea CSS CSM [16] CSM 2 BM CSM CSS AO BM X 1 (f,m) BM [22] MRI BM M (f,m) 4 M (f 1,m), M (f +1,m), M (f,m 1), M (f,m +1)) 0

7 musical noise 4 musical noise 3 musical noise 2 musical noise 1 musical noise Sources to Artifacts Ratio (SAR): 6 M (f,m) 0 M (f,m) 4 1 M (f,m) 1 MRI 2 Y (f,m) Y (f,m) = 1 2 Y (f,m) {Y (f 1,m)+Y (f +1,m) +Y (f,m 1) + Y (f,m +1)} (14) [23] Perceptual SS BM [23] ms M =3 N = f s DFT F 1 F/2 [24] Signal to Distortion Ratio (SDR): Source Image to Spatial distortion Ratio (ISR): Source to Interference Ratio (SIR): SDR ISR SIR SAR [24] SIR ISR SAR db 11 Mean Opinion Score (MOS) BM CSS 3 BM, CSM, CSS (a) CSM CSS MOS BM CSM CSS SIR BM ISR SAR BM CSM original [16] CSS β l CSS case2 CSS ISR SAR CSM original CSM CSS CSS MOS CSM 8 case2 β l CSS 3(A)

8 81 劣決定音源分離のための分離音声のケプストラムスムージング図 7(b) はケプストラムスムージング手法 5.1 節減に効果的であることが分かると同一の CSM や CSS と 5.2 節で述べたミュージこのようにケプストラムスムージング手法 CSM カルノイズ低減手法との比較結果を示している CSM や CSS はミュージカルノイズ低減に効果的であるや CSS の ISR と SAR は他手法より低いが一方でまた図 7(a) より CSS の性能とパラメータ βpea ミュージカルノイズの量に着目した MOS 値は CSM との関係を読み取ることができるすなわち βpea がや CSS の方が高く他手法よりミュージカルノイズ低大きな値の場合 case1 には MOS 値が高く ISR や SAR が低いすなわちミュージカルノイズが低減され信号歪は大きい結果が得られる一方 βpea が小さな値の場合 case2, case3 には信号歪は小さいもののミュージカルノイズが顕著となるこれは 4.2 節に述べたミュージカルノイズの軽減の程度とスペクトル微細構造の保持の程度のトレードオフを示している 5.5 考察前節で述べたとおり提案法である CSS は CSM より高い ISR や SAR を持つことが分かったまた CSS 及び CSM 法はミュージカルノイズ低減に効果的であることも示されたしかし上述したとおり CSS 及び CSM 法は BM 法図 7(a) や 5.2 節に述べたミュージカルノイズ低減手法図 7(b) と比較して ISR や SAR が低くなることからケプストラムスムージング手法ではミュージカルノイズとは異なる歪が生じることが分かった実際図 7 各歪値と MOS の比較結果図 8 に著者らが聴取したところではケプストラムスムージ CSS による分離信号のスペクトログラム

9 (A) (B) BM (C) CSM (D) CSS ISR SAR CSS case2 CSM original ISR SAR MOS CSS case2 CSM 9 BM CSM CSS

10 83 10 CSM bin BM 9 BM BM CSM 9(B) (C) CSM 0 l env CSM 10 BM 1 M min 2 10 (4) BM CSM 10 CSM (10) 3.2 CSM CSM CSS BM CSS 9 (B) (D) CSS BM (A) CSS CSM CSS CSS CSM CSS 6. BSS CSM CSS CSS CSS CSS CSM CSS NTT [ 1 ] S. Hayin, Ed., Unsupervised Adaptive Filtering, Volume I: Blind Source Separation (Wiley, New Yor, 2000). [ 2 ] A. Hyvärinen, J. Karhunen and E. Oja, Independent Component Analysis (John Wiley & Sons, New Yor, 2001). [3] Ö. Yilmaz and S. Richard, Blind separation of speech mixtures via time-frequency masing, IEEE Trans. Signal Process., 52, (2004). [ 4 ] N. Roman and D. Wang, Binaural sound segregation for multisource reverberant environments, Proc. ICASSP 2004, Vol. II, pp (2004).

84 68 2 2012 [ 5 ] S. Ricard and Ö. Yilmaz, On the W-disjoint orthogonality of speech, Proc. ICASSP 2002, Vol. 1, pp. 529 532 (2002). [ 6 ] S. Arai, H. Sawada, R. Muai and S.

Zibulevsy, Blind separation of more sources than mixtures using sparsity of their short-time Fourier transform, Proc. ICA 2000, pp. 87 92 (2000). [ 8 ] A. Jourjine, S. Ricard and Ö.

Saurai and Y. Kaneda, Sound source segregation based on estimating incident angle of each frequency component of input signals acquired by multiple microphones, Acoust. Sci. & Tech.

Acoust. Soc. Am., 114, 2236 2252 (2003). [12] S. Arai, S. Maino, A. Blin, R. Muai and H.

Kadambe, A probabilistic approach for blind source separation of underdetermined convolutive mixtures, Proc. ICASSP 2003, Vol. VI, pp. 581 584 (2003). [14] S. Arai, S. Maino, A. Blin, R. Muai and H.

11 [ 5 ] S. Ricard and Ö. Yilmaz, On the W-disjoint orthogonality of speech, Proc. ICASSP 2002, Vol. 1, pp (2002). [ 6 ] S. Arai, H. Sawada, R. Muai and S. Maino, Underdetermined blind sparse source separation for arbitrarily arranged multiple sensors, Signal Process., 77, (2007). [ 7 ] P. Bofill and M. Zibulevsy, Blind separation of more sources than mixtures using sparsity of their short-time Fourier transform, Proc. ICA 2000, pp (2000). [ 8 ] A. Jourjine, S. Ricard and Ö. Yilmaz, Blind separation of disjoint orthogonal signals: Demixing N sources from 2 mixtures, Proc. ICASSP 2000, Vol. 5, pp (2000). [ 9 ] M. Aoi, M. Oamoto, S. Aoi, H. Matsui, T. Saurai and Y. Kaneda, Sound source segregation based on estimating incident angle of each frequency component of input signals acquired by multiple microphones, Acoust. Sci. & Tech., 22, (2001). [10] S. Ricard, R. Balan and J. Rosca, Real-time time-frequency based blind source separation, Proc. ICA 2001, pp (2001). [11] N. Roman, D. Wang and G.J. Brown, Speech segregation based on sound localization, J. Acoust. Soc. Am., 114, (2003). [12] S. Arai, S. Maino, A. Blin, R. Muai and H. Sawada, Blind separation of more speech than sensors with less distortion by combining sparseness and ICA, Proc. IWAENC 2003, pp (2003). [13] J.M. Peterson and S. Kadambe, A probabilistic approach for blind source separation of underdetermined convolutive mixtures, Proc. ICASSP 2003, Vol. VI, pp (2003). [14] S. Arai, S. Maino, A. Blin, R. Muai and H. Sawada, Underdetermined blind separation for speech in real environments with sparseness and ICA, Proc. ICASSP 2004, Vol. III, pp (2004). [15] S. Arai, H. Sawada, R. Muai and S. Maino, Blind sparse source separation with spatially smoothed time-frequency masing, Proc. IWAENC 2006 (2006). [16] N. Madhu, C. Breithaupt and R. Martin, Temporal smoothing of spectral mass in the cepstral domain for speech separation, Proc. ICASSP 2008, pp (2008). [17] Y. Ansai, S. Arai, S. Maino, T. Naatani, T. Yamada, A. Naamura and N. Kitawai, Cepstral smoothing of separated signals for underdetermined speech separation, Proc. ISCAS 2010, pp (2010). [18] P. Bofill and M. Zibulevsy, Blind separataion of more sources than mixtures using sparsity of their short-time-fourier transform, Proc. ICA 2000, pp (2000). [19] A. Blin, S. Arai and S. Maino, Blind source separation when speech signals outnumber sensors using an sparseness mixing matrix combination, Proc. IWAENC 2003, pp (2003). [20] Y. Izumi, N. Ono and S. Sagayama, Sparsenessbased 2ch BSS using EM algorithm in reverberant environment, Proc. WASPAA, pp (2007). [21],,, BSS,, pp (2008). [22],,,, pp (2004.3). [23] N. Virag, Single channel speech enhancement based on masing properties of the human auditory system, IEEE Trans. Speech Audio Process., 7, (1999). [24] E. Vincent, H. Sawada, P. Bofill, S. Maino and J.P. Rosca, First stereo audio source separation evaluation campaign: Data, algorithms and results, Proc. ICA 2007, pp (2007) IEEE 56 NTT IEEE Distinguished Lecturer IEEE Fellow Fellow 03 IEEE 11 IEEE

12 ATR 12 / IEEE IEEE Fellow

IPSJ SIG Technical Report 1, Instrument Separation in Reverberant Environments Using Crystal Microphone Arrays Nobutaka ITO, 1, 2 Yu KITANO, 1

IPSJ SIG Technical Report 1, Instrument Separation in Reverberant Environments Using Crystal Microphone Arrays Nobutaka ITO, 1, 2 Yu KITANO, 1 1, 2 1 1 1 Instrument Separation in Reverberant Environments Using Crystal Microphone Arrays Nobutaka ITO, 1, 2 Yu KITANO, 1 Nobutaka ONO 1 and Shigeki SAGAYAMA 1 This paper deals with instrument separation