ComputeNoveltyFunction

computes the novelty function for onset detection
>
> supported novelty measures are:
>  'Flux',
>  'Laroche',
>  'Hainsworth'
>
> @param cNoveltyName: name of the novelty measure
> @param x: time domain sample data, dimension channels X samples
> @param f_s: sample rate of audio data
> @param afWindow: FFT window of length iBlockLength (default: hann), can be [] empty
> @param iBlockLength: internal block length (default: 4096 samples)
> @param iHopLength: internal hop length (default: 512 samples)
>
> @retval f frequency
> @retval t time stamp for the frequency value
> @retval G_T threshold function
> @retval iPeaks indices of picked onset times
 ======================================================================

0001 %computes the novelty function for onset detection
0002 %>
0003 %> supported novelty measures are:
0004 %>  'Flux',
0005 %>  'Laroche',
0006 %>  'Hainsworth'
0007 %>
0008 %> @param cNoveltyName: name of the novelty measure
0009 %> @param x: time domain sample data, dimension channels X samples
0010 %> @param f_s: sample rate of audio data
0011 %> @param afWindow: FFT window of length iBlockLength (default: hann), can be [] empty
0012 %> @param iBlockLength: internal block length (default: 4096 samples)
0013 %> @param iHopLength: internal hop length (default: 512 samples)
0014 %>
0015 %> @retval f frequency
0016 %> @retval t time stamp for the frequency value
0017 %> @retval G_T threshold function
0018 %> @retval iPeaks indices of picked onset times
0019 % ======================================================================
0020 function [d, t, G_T, iPeaks] = ComputeNoveltyFunction (cNoveltyName, x, f_s, afWindow, iBlockLength, iHopLength)
0021 
0022     % set function handle
0023     hNoveltyFunc = str2func (['Novelty' cNoveltyName]);
0024 
0025     % set default parameters if necessary
0026     if (nargin < 6)
0027         iHopLength = 512;
0028     end
0029     if (nargin < 5)
0030         iBlockLength = 4096;
0031     end
0032     if (nargin < 4 || isempty(afWindow))
0033         afWindow = hann(iBlockLength, 'periodic');
0034     end
0035 
0036     % compute FFT window function
0037     if (length(afWindow) ~= iBlockLength)
0038         error('window length mismatch');
0039     end        
0040 
0041     % parametrization of smoothing filters
0042     fSmoothLpLenInS = 0.07;
0043     fThreshLpLenInS = 0.14;
0044     iSmoothLpLen = max(2, ceil(fSmoothLpLenInS * f_s / iHopLength));
0045     iThreshLpLen = max(2, ceil(fThreshLpLenInS * f_s / iHopLength));
0046 
0047     % pre-processing: down-mixing
0048     x = ToolDownmix(x);
0049 
0050     % pre-processing: normalization (not necessary for many features)
0051     x = ToolNormalizeAudio(x);
0052 
0053     % in the real world, we would do this block by block...
0054     [X, f, t] = ComputeSpectrogram(x, ...
0055                                 f_s, ...
0056                                 afWindow, ...
0057                                 iBlockLength, ...
0058                                 iHopLength);
0059 
0060     % novelty function
0061     d = hNoveltyFunc(X, f_s);
0062     
0063     % smooth novelty function
0064     b = ones(iSmoothLpLen,1)/iSmoothLpLen;
0065     d = filtfilt (b, 1, d);
0066     d(d<0) = 0;
0067     
0068     % compute threshold
0069     iLen = min(iThreshLpLen, floor(length(d) / 3));
0070     b = ones(iLen, 1) / iLen;
0071     G_T = .4*mean(d(2:end)) + filtfilt (b, 1, d);
0072     
0073     [fPeaks, iPeaks] = findpeaks(max(0, d - G_T)); 
0074 end
0075