BSS Oracle Toolbox Version 2.1 User Guide
Contents
1. 18 bss_nearopt_monomask 19 bssomearopt binmask i e Pag m Done a eae de Seen s 20 bss mearopt pinvmask gt Go ea cy dO So Bee ye hs 21 3 4 Time frequency transforms 22 Inde ute re Goad are Whee Bee fom du Pe te Sy oe Set 22 Indet testa e tels ate A are a ds Tiana dol vr es tud CEU UE ni a 23 cond RD DEPT cat 24 cu EU cM VERE Te ety NC Re Ug PEOR the na vie sate a Ae ied et ad etas ee ied coge eios 25 3 5 Filtering and masking functions 26 apply cmultifitt temp i2 uo RW RUE poem EROR xx Reds ue ety ter dern 26 apply multifilt freg od hae mot a ed ede Eddie go e d 27 apply pinvmask inst 4 ew nee ae a qn eue Se a ae a 28 apply_pinvmask_conv 29 3 6 Aiixiliary functions s s 050 p63 Wis sce Ps doe 30 optim coeffs x od eek Dia Piny filt 52 33x a eU Gee Sdn Sdr d ev cA mi egi id E tg 4 Example data and applications 4 1 Sources and filters 42 Seripts cog nee Ba das eR e eu ee eS Bibliography CONTENTS Chapter 1 Getting started 1 1 Download and install Two versions of the BSS Oracle toolbox are available The basic version of the toolbox which includes the main Matlab programs and this u2. Example data and applications The full version of BSS Oracle contains example sources filters and scripts that were used to plot the figures of the reference publications 4 1 Sources and filters data mixk sj wav data ir_t mat data ir movei 250ms mat data ir move2 250ms mat data ir move3 250ms mat See 3 Section 3 4 2 Scripts multifilti multifilt2 multifilt3 multifilt4 monomaski m m m m m monomask2 m monomask3 m multimaski m multimask2 m bbasis monomaski m bbasis monomask2 m bbasis monomask3 m robusti m robust2 m source j of mixture k with 1 j lt 3 music for 1 k lt 10 speech for 11 k lt 20 mixing impulse responses with reverberation time t equal to anechoic 50ms 250ms or 1 25s with source 1 at 40 mixing impulse responses with 250 ms reverberation time with source 1 at 38 mixing impulse responses with 250 ms reverberation time with source 1 at 36 mixing impulse responses with 250 ms reverberation time with source 1 at 32 plots figure 2 of plots figure 3 of plots figure 4 of plots figure 4 of plots figure 5 of plots figure 7 of plots figure 6 of plots figure 7 of plots figure 8 of plots figure 11 of 2 plots figure 12 of 2 plots figures 13 and 14 of 2 plots figure 9 of 3 plots figure 10 of 3 w Q2 Q9 ND Q2 CO ND C2 C2 33 34 CHAPTER 4 EXAMPLE DATA AND APPLICATIONS Bibliography 1 E Vincent and R Gribonval Blind criterion
3. X S Dmin Dmax pcos Outputs Se btree SDR stree Reference See 1 I x matrix containing the multichannel mixture signal J x T matrix containing the target signals source images minimal packet depth maximal packet depth true for CP basis with sine window default false for WP basis with symmlet 8 J x T matrix containing the oracle estimates of the target signals truncated to the same time range as the original signals 1 x 2Pmaxtl 1 vector of binary values representing the tree structure corre sponding to the oracle best basis 2N x Dmax Dmin 1 x J table containing the oracle masking coefficients for each scale with N nextpow2 T achieved SDR in deciBels before truncation of the target estimates 1 x 2Pmax 1 vector containing the oracle distortion for all basis elements infinite for disallowed scales 16 CHAPTER 3 REFERENCE MANUAL bss_oracle_bbasis_pinvmask Oracle estimator for multichannel source separation of instantaneous mixtures by time frequency masking and mixing matrix pseudo inversion using the best CP WP basis Syntax Se btree W SDR stree bss_oracle_bbasis_pinvmask X S A Dmin Dmax Se btree W SDR stree bss oracle bbasis pinvmask X S A Dmin Dmax Ja Se btree W SDR stree bss oracle bbasis pinvmask X S A Dmin Dmax Ja pcos Inputs X I x matrix containing the multichannel mixture signal S J x T matrix containing the target signals so
4. TIME FREQUENCY TRANSFORMS 25 istft Inverse Short Term Fourier Transform using a sine window Syntax x istft X x istft X M Inputs 4 1 matrix containing a set of STFT coefficients for positive frequencies M step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 4 default 4 Output x 1x NM vector containing the inverse STFT signal If x is a signal of length T X stft x L and y istft X then x y 1 T 26 CHAPTER 3 REFERENCE MANUAL 3 5 Filtering and masking functions apply_multifilt_temp Apply time domain demixing filters Syntax Se apply multifilt temp X W Inputs I x matrix containing the multichannel mixture signal W J x I x L table containing the coefficients of the demixing filters delays from 41 to 2 2 2 Output Se J x T L 1 matrix containing the demixed signals Reference See 3 Section 4 1 3 5 FILTERING AND MASKING FUNCTIONS 27 apply_multifilt_freq Apply frequency domain demixing matrices using STFT with sine window Syntax Se apply multifilt freq X W Se apply_multifilt_freq X W M Inputs X I x T matrix containing the multichannel mixture signal W JxlIx 4 1 table containing complex demixing matrices for positive fre quencies with L being the length of the STFT window in samples M step between successive windows in samples must be a multiple of 2 a divid
5. and oracle bound for instantaneous audio source separation using adaptive time frequency representations In Proc IEEE Workshop on Applications of Signal Processing to Audio and Acoustics WASPAA 2007 2 E Vincent R Gribonval and M D Plumbley Oracle estimators for the benchmarking of source separation algorithms Technical Report C4DM TR 06 03 Queen Mary University of London 2006 3 E Vincent R Gribonval and M D Plumbley Oracle estimators for the benchmarking of source separation algorithms Signal Processing 87 8 1933 1950 2007 35
6. BSS Oracle Toolbox Version 2 1 User Guide Emmanuel VINCENT R mi GRIBONVAL Mark D PLUMBLEY October 12 2007 Contents 1 Getting started 5 1 1 Download and install 5 1 2 Software dependencies 5 1 9 Gettimp sa em di a gl ee Ses ee Ga MP ie ad 5 LA Citation b er bee Bee le le Gap ede DURS APE ttes 6 1 5 Lic nse S ouo Reo mA es Que UR ERE EUER Seo 6 2 Content 7 2 1 Reference publications 7 2 Principles Ppa oes uo dos ce ee ese Seb E 7 2 9 Summary of the content a rare e a 2 oh e depen ee ee a 7 3 Reference manual 9 3 1 General notations us 25 x Ross ee ee AU ee d 9 3 2 Oracle source estimators 10 bss oracle m ltifilt sage ba cee de nl Et p UR ees 10 bss_ oracle monomask 11 bsssoracle binmask 123 3 x44 D ELDER Mint e m ROLE 12 bss oracle pinvHask 24 014 x3 379 RO Ro f bU douanes MUS n 13 bss oracle bbasis monomask 14 bss oracle bbasis binmask 15 bss oracle bbasis pinvmask 16 bss_oracle gbasis monomask 17 3 3 Near optimal source estimators 18 bss_nearopt_multifilt
7. R stree bss oracle bbasis monomask x S Dmin Dmax pcos mreal mconst Inputs x 1 x T vector containing the single channel mixture signal 5 J x T matrix containing the target signals e g source images Dmin minimal packet depth Dmax maximal packet depth pcos true for CP basis with sine window default false for WP basis with symmlet 8 mreal true for real valued masking default false for binary masking mconst true when the masks are subject to a unitary sum constraint default false otherwise Outputs Se J x T matrix containing the oracle estimates of the target signals truncated to the same time range as the original signals btree 1 x 2Pmaxtl 1 vector of binary values representing the tree structure corre sponding to the oracle best basis W 2N x Dmax Dmin 1 x J table containing the oracle masking coefficients for each scale with N nextpow2 T SDR achieved SDR in deciBels before truncation of the target estimates stree 1 x 2Pmaxtl 1 vector containing the oracle distortion for all basis elements infinite for disallowed scales Reference See 2 Section 7 2 3 2 ORACLE SOURCE ESTIMATORS 15 bss_oracle_bbasis_binmask Oracle estimator for multichannel source separation by constrained binary time frequency mask ing using the best CP WP basis Syntax Se btree W SDR stree bss oracle bbasis binmask X S Dmin Dmax Se btree W SDR stree bss oracle bbasis binmask X S Dmin Dmax pcos Inputs
8. e containing the derived source images Reference See 3 Section 6 4 30 CHAPTER 3 REFERENCE MANUAL 3 6 Auxiliary functions optim_coeffs Oracle constrained real valued masking coefficients for a single basis element Syntax wo disto optim_coeffs r Input r J x 1 vector containing ratios between MDCT CP or WP coefficients or real parts of ratios between STFT coefficients of the targets and the mixture for a single basis element Outputs wo oracle masking coefficients for this basis element disto achieved distortion Reference See 3 Section 5 2 3 6 AUXILIARY FUNCTIONS 31 pinv_filt Pseudo inversion of a filter system Syntax W B SIR pinv filt A zdel L Inputs A Ix JxT table containing filters of length delays from zdel 1 to T zdel zdel sample index corresponding to zero delay L length of the pseudo inverse filters in samples Outputs W J x I x L table containing pseudo inverse filters delays from L 1 to 5 J x J x T L 1 product of W and A delays from zdel 2 to zdel 5 SIR achieved SIR in deciBels Reference See 2 Section 4 3 3 32 CHAPTER 3 REFERENCE MANUAL sdr Signal to Distortion Ratio Syntax SDR sdr Se S Inputs Se J x T matrix containing the estimated signals 5 J x T matrix containing the target signals Output SDR achieved SDR in deciBels Reference See 3 Section 2 3 Chapter 4
9. er of L and smaller than 5 default 4 Output Se J x NM matrix containing the demixed signals with N ceil 5 See 3 Section 4 4 28 CHAPTER 3 REFERENCE MANUAL apply_pinvmask_inst Apply multichannel time frequency masks with mixing matrix pseudo inversion using MDCT with sine window Syntax Se apply_pinvmask_inst X A W Inputs X A W Output Se Reference I x matrix containing the multichannel mixture signal I x J real valued mixing matrix 5 N J table of binary coefficients indicating the source activity patterns with L being the length of the MDCT window in samples and N ceil 27 2 Ay x J table containing the derived source images See 3 Section 6 1 2 3 5 FILTERING AND MASKING FUNCTIONS 29 apply_pinvmask_conv Apply multichannel time frequency masks with mixing matrix pseudo inversion using STFT with sine window Syntax Se apply_pinvmask_conv X A W Se apply pinvmask conv X A W M Inputs X I x T matrix containing the multichannel mixture signal A IxJx 4 1 table containing complex mixing matrices for positive frequen cies with L being the length of the STFT window in samples W L x N x J table of binary coefficients indicating the source activity patterns with N ceil 27 M step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 5 default 5 Output Se J x NM x J tabl
10. le containing the near optimal estimates of the target signals truncated to the same time range as the original signals 4 1 x N x J table containing near optimal masks with N ceil i achieved SDR in deciBels before truncation of the target estimates See 3 Section 6 4 3 3 NEAR OPTIMAL SOURCE ESTIMATORS 21 bss_nearopt_pinvmask Near optimal estimator for multichannel source separation of possibly convolutive mixtures by time frequency masking and mixing matrix pseudo inversion using STFT with sine window activity patterns derived for each time frequency point separately Syntax Se W SDR bss_nearopt_pinvmask X S A Se W SDR bss_nearopt_pinvmask X S A Ja Se W SDR bss_nearopt_pinvmask X S A Ja M Inputs X I x matrix containing the multichannel mixture signal S I x T x J table containing the target signals source images A IxJx 4 1 table containing complex mixing matrices for positive frequencies may be different from the ones actually used to generate S with L being the length of the STFT window in samples Ja number of active sources per time frequency point by default or if Ja 0 the best number is estimated for each time frequency point M step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 4 default 4 Outputs Se I x T x J table containing the near optimal estimates of the target signals truncated to the same time range a
11. mator for multichannel source separation of instantaneous mixtures by time frequency masking and mixing matrix pseudo inversion using MDCT with sine window Syntax Se W SDR bss_oracle_pinvmask X S L A Se W SDR bss oracle pinvmask X S L A Ja Inputs X I x matrix containing the multichannel mixture signal 5 I x T x J table containing the target signals source images L length of the MDCT window in samples must be a multiple of 4 A I x J real valued mixing matrix may be different from the one actually used to generate S Ja number of active sources per time frequency point by default or if Ja 0 the best number is estimated for each time frequency point Outputs Se I x T x J table containing the oracle estimates of the target signals truncated to the same time range as the original signals W L x N x J table of binary coefficients indicating the oracle source activity patterns with N ceil 27 SDR achieved SDR in deciBels before truncation of the target estimates Reference See 3 Section 6 2 14 CHAPTER 3 REFERENCE MANUAL bss_oracle_bbasis_monomask Oracle estimator for single channel source separation by time frequency masking using the best CP WP basis Syntax Se btree W SDR stree bss oracle bbasis monomask x S Dmin Dmax Se btree W SDR stree bss oracle bbasis monomask x S Dmin Dmax pcos Se btree W SDR stree bss oracle monomask x S Dmin Dmax pcos mreal Se btree W SD
12. nal 5 J x T matrix containing the target signals e g source images L length of the STFT window in samples must be a multiple of 4 M step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 4 default 4 mreal true for real valued masking default false for binary masking mconst true when the masks are subject to a unitary sum constraint default false otherwise Outputs Se J x T matrix containing the near optimal estimates of the target signals trun cated to the same time range as the original signals W 5 1 N x J table containing near optimal masks with N ceil T SDR achieved SDR in deciBels before truncation of the target estimates Reference See 3 Section 5 4 20 CHAPTER 3 REFERENCE MANUAL bss_nearopt_binmask Near optimal constrained binary time frequency masks for multichannel source separation using STFT with sine window coefficients derived for each time frequency point separately Syntax Se W SDR bss_nearopt_binmask X S L Se W SDR bss nearopt binmask X S L M Inputs Ww Outputs se W SDR Reference I x matrix containing the multichannel mixture signal I x T x J table containing the target signals source images length of the STFT window in samples must be a multiple of 4 step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 4 default 4 I x T x J tab
13. oracle gbasis monomask x S Dmin Dmax pcos mreal mconst Warnings Despite the default setting this function is mostly relevant for WP bases Due to huge memory requirements the optimal masks are not output The function stores temporary data in a temporary directory defined by the variable tmpdir default is tmp Inputs x 1x T x table containing K single channel mixture signals 5 J x T x table containing K sets of target signals e g source images Dmin minimal packet depth Dmax maximal packet depth pcos true for CP basis with sine window default false for WP basis with symmlet 8 mreal true for real valued masking default false for binary masking mconst true when the masks are subject to a unitary sum constraint default false otherwise Outputs Se J x T x K table containing the oracle estimates of the target signals truncated to the same time range as the original signals btree 1 x 2Pmaxtl 1 vector of binary values representing the tree structure corre sponding to the oracle best generic basis SDR K x 1 vector containing the achieved SDR in deciBels for each mixture before truncation of the target estimates gSDR total SDR for all mixtures stree 1 x 2Pmaxtl 1 vector containing the oracle distortion for all basis elements infinite for disallowed scales Reference See 2 Section 7 3 2 18 CHAPTER 3 REFERENCE MANUAL 3 3 Near optimal source estimators bss_nearopt_multifilt Nea
14. r optimal demixing matrices for source separation by frequency domain multichannel time invariant filtering using STFT with sine window coefficients derived for each frequency bin separately Syntax Se W SDR bss nearopt multifilt X S L Se W SDR bss nearopt multifilt X S L M Inputs X I x T matrix containing the multichannel mixture signal 5 J x T matrix containing the target signals e g sources or source images L length of the STFT window in samples must be a multiple of 4 M step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 4 default 4 Outputs Se J x T matrix containing the near optimal estimates of the target signals trun cated to the same time range as the original signals W 4 1 table containing near optimal demixing matrices for positive frequencies SDR achieved SDR in deciBels before truncation of the target estimates Reference See 3 Section 4 4 3 3 NEAR OPTIMAL SOURCE ESTIMATORS 19 bss_nearopt_monomask Near optimal time frequency masks for single channel source separation using STFT with sine window coefficients derived for each time frequency point separately Syntax Se W SDR bss_nearopt_monomask x S L Se W SDR bss_nearopt_monomask x S L M Se W SDR bss_nearopt_monomask x S L M mreal Se W SDR bss_nearopt_monomask x S L M mreal mconst Inputs x 1 x T vector containing the single channel mixture sig
15. s the original signals W 4 1 x N x J table of binary coefficients indicating the oracle source activity patterns with N ceil i SDR achieved SDR in deciBels before truncation of the target estimates Reference See 3 Section 6 4 22 CHAPTER 3 REFERENCE MANUAL 3 4 Time frequency transforms mdct Modified Discrete Cosine Transform using a sine window Syntax X mdct x L Inputs x 1 x T vector containing a single channel signal L length of the MDCT window in samples must be a multiple of 4 Output X 5 matrix containing MDCT coefficients with N ceil 27 3 4 TIME FREQUENCY TRANSFORMS 23 imdct Inverse Modified Discrete Cosine Transform using a sine window Syntax x imdct X Input X L x N matrix containing a set of MDCT coefficients Outputs x 1x NE vector containing the inverse MDCT signal If x is a signal of length T X mdct x L and y imdct X then x y 1 T 24 CHAPTER 3 REFERENCE MANUAL stft Short Term Fourier Transform using a sine window Syntax X stft x L X stft x L M Inputs x 1 x T vector containing a single channel signal L length of the STFT window in samples must be a multiple of 4 M step between successive windows in samples must be a multiple of 2 a divider of L and smaller than 5 default 5 Output X 4 1 x N matrix containing the STFT coefficients for positive frequencies with N ceil i 3 4
16. samples number of channels of the mixture signal number of source or target signals length of the demixing filters or MDCT STFT length stepsize between successive STFT windows number of mixture signals for generic oracle bases 10 CHAPTER 3 REFERENCE MANUAL 3 2 Oracle source estimators bss_oracle_multifilt Oracle estimator for source separation by multichannel time invariant filtering in the time do main Syntax Se W SDR bss_oracle_multifilt X S L Inputs X 5 L Outputs Se W SDR Reference I x matrix containing the multichannel mixture signal J x T matrix containing the target signals e g sources or source images length of the demixing filters in samples J x T matrix containing the oracle estimates of the target signals truncated to the same time range as the original signals J x I x table containing the coefficients of the oracle demixing filters delays from 1 to 2 achieved SDR in deciBels before truncation of the target estimates See 3 Section 4 2 3 2 ORACLE SOURCE ESTIMATORS 11 bss_oracle_monomask Oracle estimator for single channel source separation by time frequency masking using MDCT with sine window Syntax Se W SDR bss oracle monomask x S L Se W SDR bss oracle monomask x S L mreal Se W SDR bss oracle monomask x S L mreal mconst Inputs x 5 L mreal mconst Outputs Se W SDR Reference 1
17. ser guide can be downloaded at http bass db gforge inria fr bss_oracle bss_oracle_basic zip After unzipping this file you should get a directory called bss oracle 2 1 To install simply add the full path to this directory to your Matlab path using the command pathtool The full version of the toolbox which includes additional example data and Matlab programs can be downloaded at http bass db gforge inria fr bss_oracle bss_oracle_full zip After unzipping this file you should get in addition to the main directory called bss_oracle_2 1 two sub directories called examples and data To install simply add the full paths to the main directory and to the sub directories to your Matlab path 1 2 Software dependencies BSS Oracle consists in a set of Matlab functions and as such needs Matlab to run Some functions of BSS Oracle involving MDCT CP or WP transforms also depend on the Wavelab toolbox version 802 by D Donoho M R Duncan X Huo and O Levi available at http www stat stanford edu wavelab Follow the provided documentation for install instructions Note that Wavelab is copyrighted and cannot be redistributed together with BSS Oracle 1 8 Getting help Within Matlab you can get basic help about the toolbox by typing help bss oracle 2 1 1Matlab is a registered trademark of The MathWorks Inc 5 6 CHAPTER 1 GETTING STARTED 1 4 Citation If you use the BSS Oracle toolbox in a work that you wi
18. sh to publish please cite it as E Vincent R Gribonval and M D Plumbley BSS Oracle Toolbox Version 2 1 http bass db gforge inria fr bss_oracle 1 5 Licenses The files contained in the BSS Oracle toolbox are distributed under different licenses Therefore it is crucial that you understand which license applies to each file before attempting to redis tribute or modify some files The files contained in the main directory bss_oracle_2 1 and in the subdirectory examples are distributed under the terms of the GNU General Public License GPL version 2 A copy of the GPL is distributed along with the toolbox in the file LICENSE txt The music sound files contained in the subdirectory data are distributed under specific Cre ative Commons licenses For more details about the license applying to each file see the file data LICENSES txt All other files of the subdirectory data are license free Chapter 2 Content The purpose of the BSS Oracle toolbox is to compute the best performance achievable by a class of source separation algorithms in an evaluation framework where target signals are known It does not provide any blind source separation method 2 1 Reference publications The mathematical details underlying the toolbox are described in 3 1 2 2 2 Principle Let us suppose that we observe a mixture signal x t from which we want to extract a set of source signals y t Within a given class of source separation algori
19. thms the estimated signal t can always be expressed under the form F f x 0 with 6 6 2 1 where f is a fixed parametric function 0 a vector of separation parameters and a set of acceptable parameters Different algorithms correspond to different ways of estimating 0 Assuming that the target signal y t is known the separation performance of a given algo rithm can be evaluated using the Euclidean distortion measure d y y lly yll 2 2 The oracle estimator of the target signal is then defined by yly X f x X 9 2 3 where 0 x O is the set of parameters resulting in the smallest distortion among the set of acceptable parameters O x O arg min d f x 0 y 2 4 2 3 Summary of the content The basic version of the toolbox implements oracle source estimators for four classes of algo rithms multichannel time invariant filtering single channel time frequency masking multichan nel time frequency masking and best basis masking In some cases the exact oracle estimators 7 8 CHAPTER 2 CONTENT cannot be computed due to high memory and or computational time requirements Thus near optimal source estimators are implemented instead The full version of the toolbox also contains example data and routines that were used to create the figures of the reference publications Chapter 3 Reference manual 3 1 General notations length of the signals
20. urce images A I x J real valued mixing matrix may be different from the one actually used to generate S Dmin minimal packet depth Dmax maximal packet depth Ja number of active sources per time frequency point by default or if Ja 0 the best number is estimated for each time frequency point pcos true for CP basis with sine window default false for WP basis with symmlet 8 Outputs Se J x T matrix containing the oracle estimates of the target signals truncated to the same time range as the original signals btree 1x QP max 1 1 vector of binary values representing the tree structure corre sponding to the oracle best basis W 2N x Dmax Dmin 1 x J table containing the oracle masking coefficients for each scale with N nextpow2 T SDR achieved SDR in deciBels before truncation of the target estimates stree 1 x 2Pmaxtl _ 1 vector containing the oracle distortion for all basis elements infinite for disallowed scales Reference See 1 Section 3 2 3 2 ORACLE SOURCE ESTIMATORS 17 bss_oracle_gbasis_monomask Oracle estimator for single channel source separation of several mixtures by time frequency masking using the best generic CP WP basis Syntax Se btree SDR gSDR stree bss_oracle_gbasis_monomask x S Dmin Dmax Se btree SDR gSDR stree bss oracle gbasis monomask x S Dmin Dmax pcos Se btree SDR gSDR stree bss_oracle_gbasis_monomask x S Dmin Dmax pcos mreal Se btree SDR gSDR stree bss
21. x T vector containing the single channel mixture signal J x T matrix containing the target signals e g source images length of the MDCT window in samples must be a multiple of 4 true for real valued masking default false for binary masking true when the masks are subject to a unitary sum constraint default false otherwise J x T matrix containing the oracle estimates of the target signals truncated to the same time range as the original signals L x N x J table containing the oracle masks with N ceil 22 achieved SDR in deciBels before truncation of the target estimates See 3 Section 5 2 12 CHAPTER 3 REFERENCE MANUAL bss_oracle_binmask Oracle estimator for multichannel source separation by constrained binary time frequency mask ing using MDCT with sine window Syntax Se W SDR bss_oracle_binmask X S L Inputs X I x matrix containing the multichannel mixture signal 5 I x T x J table containing the target signals source images 1 length of the MDCT window in samples must be a multiple of 4 Outputs Se I x T x J table containing the oracle estimates of the target signals truncated to the same time range as the original signals W L x N x J table containing the oracle masks with N ceil 22 SDR achieved SDR in deciBels before truncation of the target estimates Reference See 3 Section 6 2 3 2 ORACLE SOURCE ESTIMATORS 13 bss_oracle_pinvmask Oracle esti
Download Pdf Manuals
Related Search