covarep / covarep Goto Github PK

Currently the frequency check only tests if sample rate is above 16kHz and then down-samples the audio to 16kHz. I suggest to change this to capture frequencies that are "not equal" to 16000 and change them.

Line 77: if fs~=16000

Add CPP and its time/quefrency smoothing variants to the glottalsource code

recently found an issue with formant_CGDZP when it breaks at:

120 formantPeaks=[formantPeaks ; peakIndex];

and

170 while(possibleValues(1)==0) %discard zero entries

at 120 the crash happens due to mismatching when concatenating
at 170 the crash happens due to access attempt in empty array.

i made a fix at work and will commit soon when double checked that it works.
cheers

input: add zeros (maybe about 500 are sufficient) at beginning of file and it will break

From Chin Chii Yeh: "I think that line 99 and 100 should be placed above line 93, so that the size of win and ccp and cc will be the same when min(order, dftlen/2-1)=dftlen/2-1. If not there might be error in line 128."

In particular in the glottalsource files, a lot of the code could be made more efficient by making it more vectorised.

@ThomasDrugman, can we clip the values directly inside the pitch_srh.m code ?
Or this behavior was expected by you ?

See discussion at fb3a151.

When loading system_net_creak in octave, the following errors pop up:

warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nnweight\dotprod.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nnweight\dotprod.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nnweight\dotprod.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nnnetinput\netsum.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nnnetinput\netsum.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nntransfer\tansig.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nntransfer\logsig.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nntransfer\tansig.m
warning: skipping over ''
warning: load: can't find the file C:\Program Files\MATLAB\R2011a\toolbox\nnet\nnet\nntransfer\logsig.m
warning: skipping over ''
warning: no constructor for class network
warning: load: element has been converted to a structure

Since system_net_creak isn't successfully loaded, creaky_voice_do_detection fails at line 87

> creak_pp=sim(net,X);
error: Structure doesn't seem to be a neural network

I'm trying to extract the COVAREP feature from a database at 8KHz. The files were collected exactly the same way, however I'm facing some problems running the pitch_srh and gci_sedreams of methods for a few of them (less than 5% I would say).

In pitch_srh I get the following error:
Attempted to access Spec(4005); index out of bounds because numel(Spec)=4000.

Error in pitch_srh>SRH_EstimatePitch (line 154)
SRHs(freq)=(Spec(freq)+Spec(2_freq)+Spec(3_freq)+Spec(4_freq)+Spec(5_freq))-(Spec(round(1.5_freq))+Spec(round(2.5_freq))+Spec(round(3.5_freq))+Spec(round(4.5_freq)));

Error in pitch_srh (line 92)
[F0s,SRHVal,time] = SRH_EstimatePitch(res',fs,f0min,f0max,hopsize);

Error in COVAREP_feature_extraction (line 107)
[srh_f0,srh_vuv,~,srh_time] = pitch_srh(x,fs,F0min,F0max, ...

In gci_sdreams I get the following error:
Improper assignment with rectangular empty matrix.

Error in gci_sedreams (line 179)
gci(Ind)=start+posi-1;

Error in COVAREP_feature_extraction (line 117)
GCI = gci_sedreams(x,fs,F0med,1); % SEDREAMS GCI detection

Do you think this could have anything to do with the 8K sampling frequency?

Thanks!
José

memory goes out of reasonable bounds.
i will fix it with a temporary workaround:

specMat = zeros(fs, size(frameMatWinMean,2));
for i = 1:size(frameMatWinMean,2)
specMat(:,i) = abs( fft(frameMatWinMean(:,i),fs) )';
end

in addition i will add clear commands to delete unnecessary matrices whenever possible.

There is a findpeaks inbuilt function in matlab and also one in VOICEBOX. Both functions output peak amplitude and index location but in different order. I propose renaming the one in VOICEBOX findpeaks_VB.m to clear this up

Function COVAREP_feature_extraction.m, line 160:
MCEP(m,:) = spec2mfcc(hspec2spec(Ete), fs, MCEP_ord)';
The function 'spec2mfcc' is not defined anymore. I think I solved the problem using:
MCEP(m,:) = hspec2fwcep(Ete, fs, MCEP_ord)';

To test if the methods results do not vary from one version to the next.

I've been feeding get_kd_creak_features() individual audio files of vowels, segmented out of a larger recording. I've found that with some audio files with duration less than 100 ms, pitch_srh() returns an index error. I've put up a sample audio file that has this problem: http://jofrhwld.github.io/assets/test_11000.wav

> [x,fs] = wavread("test_11000.wav");
> [H2H1,res_p,ZCR,F0,F0mean,enerN,pow_std,creakF0] = get_kd_creak_features(x,fs);
error: SRH_EstimatePitch: A(I): index out of bounds; value 1 out of bound 0
error: called from:
error:   covarep/glottalsource/pitch_srh.m at line 171, column 10
error:   covarep/glottalsource/pitch_srh.m at line 87, column 24
error:   covarep/glottalsource/creaky_voice_detection/private/get_kd_creak_features.m at line 73, column 11

I wasn't sure if this was an issue with duration or samples, and I got the same error on the same bit of audio sampled at 16000.

Add a small set of "annoying" recordings and run all methods on them systematically to check the methods are robust enough (do not crash easily).

I suggest to add one on the website in order to advertise the project.
A text-version duplicate of this list could be present in the doc directory also.