This question relates to SciPy's Short-time Fourier Transform function for signal processing.
For some reason I don't understand, the size of the output 'array of sample frequencies' is exactly equal to the hop size. From the documentation:
nperseg : int, optional
Length of each segment. Defaults to 256.
noverlap : int, optional
Number of points to overlap between segments. If None, noverlap = nperseg // 2. Defaults to None. When specified, the COLA constraint must be met (see Notes below).
f : ndarray
Array of sample frequencies.
hop size H = nperseg - noverlap
I'm new to signal processing and Fourier transforms, but as far as I understand a STFT is just chopping an audio file into segments ('time frames') on which you perform a Fourier transform. So if I want to do a STFT on 100 time frames, I'd expect the output to be a matrix of size 100 x F, where F is an array of measured frequencies ('measured' probably isn't the right word here but you know what I mean).
This is kinda what SciPy's implementation does, but the size of f here is what bothers me. It's supposed to be an array describing the different frequencies, like [0Hz 500Hz 1000Hz], and it does, but for some reasons its size exactly the same as the hop size. If the hop size is 700, the number of measured frequencies is 700.
The hop size is the number of samples (i.e. time) between each time frame, and is correctly calculated as H = nperseg - noverlap, but what does this have to do with the frequency array?
Edit: Related to this question
An FFT is an square matrix transform from one orthogonal basis to another of the same dimension. This is because N is the exact number of orthogonal (e.g. that don't interfere with one another) complex sinusoids that fit in a time domain vector of length N.
A longer time vector can contain more frequency information (e.g. it's hard to tell 2 frequencies apart using just 3 sample points, but much easier with 3000 samples, etc.)
You can zero-pad your short time vector of length N to use a longer FFT, but that is identical to interpolating a nice curve between N frequency points, which makes all the FFT results interdependent.
For many purposes (visualization, etc.) an STFT is overlapped, where the adjacent segments share some overlapped data instead of just being end-to-end. This gives better time locality (e.g. the segments can be spaced closer but still be long enough so that each one can provide the frequency resolution required).
i have 4 (2 Rates and 2 Times) signals and i need to equalize their sizes. I cut them first off, because i need that too. Size of first time and rate signal is 3901, another 830. But not just to remove elements, i want to keep the curve. I thought i need interpolation and tried "resample" but it is not perfect. Looks like photo. How should i improve my codes? Any idea?
index=time >= 9.6 & tsyn <= 13.5; %time boundaries of first time signal
time1=tsyn(index); %first time signal
time_f=resample(time1,830,3901);
Rate1=CLU_YR1(index) %first rate signal
Rate_f=resample(Rate1,830,3901);
index2 = cm.Time.data >= 26.3 & cm.Time.data <= 30.45; %time boundaries of second time signal
time2=cm.Time.data(index2) %second time signal
Rat2=cm.BodySensor_SC1_Omega_B_z.data*(-180/pi) %second rate signal
Rate_p=Rat2(index2)
I suppose some of the misfit of your curves originates from the fact that the original sequence is not (close to) 0 at the end of the vector. From the matlab resample documentation:
When filtering, resample assumes that the input sequence, x, is zero before and after the samples it is given. Large deviations from zero at the endpoints of x can result in unexpected values for y.
What the best alternative is depends on what it is you want to do next. If you want to have a new, upsampled version (i.e. both signals of length 3901), you could look at interp1, which supports several different methods. If you choose to do this, keep in mind that the values in y(t) will be interpolated according to the values of t you provide. Since your time arrays do not seem to be aligned (one is between 9.6 and 13 sec, the other between 26 and 30), you are probably best off doing something along the lines of:
y_new = interp1( linspace(1,100,830), rate_p, linspace(1,100,3901), 'linear');
and the same for the time array.
I am trying to implement a frequency domain phase shift but there are few points on which I am not sure.
1- I am able to get a perfect reconstruction from a sine or sweep signal using a hanning window with a hop size of 50%. Nevertheless, how should I normalise my result when using a hop size > 50%?
2- When shifting the phase of low frequency signals (f<100, window size<1024, fs=44100) I can clearly see some non-linearity in my result. Is this because of the window size being to short for low frequencies?
Thank you very much for your help.
clear
freq=500;
fs=44100;
endTime=0.02;
t = 1/fs:1/fs:(endTime);
f1=linspace(freq,freq,fs*endTime);
x = sin(2*pi*f1.*t);
targetLength=numel(x);
L=1024;
w=hanning(L);
H=L*.50;% Hopsize of 50%
N=1024;
%match input length with window length
x=[zeros(L,1);x';zeros(L+mod(length(x),H),1)];
pend=length(x)- L ;
pin=0;
count=1;
X=zeros(N,1);
buffer0pad= zeros(N,1);
outBuffer0pad= zeros(L,1);
y=zeros(length(x),1);
delay=-.00001;
df = fs/N;
f= -fs/2:df:fs/2 - df;
while pin<pend
buffer = x(pin+1:pin+L).*w;
%append zero padding in the middle
buffer0pad(1:(L)/2)=buffer((L)/2+1: L);
buffer0pad(N-(L)/2+1:N)=buffer(1:(L)/2);
X = fft(buffer0pad,N);
% Phase modification
X = abs(X).*exp(1i*(angle(X))-(1i*2*pi*f'*delay));
outBuffer=real(ifft(X,N));
% undo zero padding----------------------
outBuffer0pad(1:L/2)=outBuffer(N-(L/2-1): N);
outBuffer0pad(L/2+1:L)=outBuffer(1:(L)/2);
%Overlap-add
y(pin+1:pin+L) = y(pin+1:pin+L) + outBuffer0pad;
pin=pin+H;
count=count+1;
end
%match output length with original input length
output=y(L+1:numel(y)-(L+mod(targetLength,H)));
figure(2)
plot(t,x(L+1:numel(x)-(L+mod(targetLength,H))))
hold on
plot(t,output)
hold off
Anything below 100 Hz has less than two cycles in your FFT window. Note that a DFT or FFT represents any waveform, including a single non-integer-periodic sinusoid, by possibly summing up of a whole bunch of sinusoids of very different frequencies. e.g. a lot more than just one. That's just how the math works.
For a von Hann window containing less than 2 cycles, these are often a bunch of mostly completely different frequencies (possibly very far away in terms of percentage from your low frequency). Shifting the phase of all those completely different frequencies may or may not shift your windowed low frequency sinusoid by the desired amount (depending on how different in frequency your signal is from being integer-periodic).
Also for low frequencies, the complex conjugate mirror needs to be shifted in the opposite direction in phase in order to still represent a completely real result. So you end up mixing 2 overlapped and opposite phase changes, which again is mostly a problem if the low frequency signal is far from being integer periodic in the DFT aperture.
Using a longer window in time and samples allows more cycles of a given frequency to fit inside it (thus possibly needing a lesser power of very different frequency sinusoids to be summed up in order to compose, make up or synthesize your low frequency sinusoid); and the complex conjugate is farther away in terms of FFT result bin index, thus reducing interference.
A sequence using any hop of a von Hann window that in 50% / (some-integer) in length is non-lossy (except for the very first or last window). All other hop sizes modulate or destroy information, and thus can't be normalized by a constant for reconstruction.
I have two sensors seperated by some distance which receive a signal from a source. The signal in its pure form is a sine wave at a frequency of 17kHz. I want to estimate the TDOA between the two sensors. I am using crosscorrelation and below is my code
x1; % signal as recieved by sensor1
x2; % signal as recieved by sensor2
len = length(x1);
nfft = 2^nextpow2(2*len-1);
X1 = fft(x1);
X2 = fft(x2);
X = X1.*conj(X2);
m = ifft(X);
r = [m(end-len+1) m(1:len)];
[a,i] = max(r);
td = i - length(r)/2;
I am filtering my signals x1 and x2 by removing all frequencies below 17kHz.
I am having two problems with the above code:
1. With the sensors and source at the same place, I am getting different values of 'td' at each time. I am not sure what is wrong. Is it because of the noise? If so can anyone please provide a solution? I have read many papers and went through other questions on stackoverflow so please answer with code along with theory instead of just stating the theory.
2. The value of 'td' is sometimes not matching with the delay as calculated using xcorr. What am i doing wrong? Below is my code for td using xcorr
[xc,lags] = xcorr(x1,x2);
[m,i] = max(xc);
td = lags(i);
One problem you might have is the fact that you only use a single frequency. At f = 17 kHz, and an estimated speed-of-sound v = 340 m/s (I assume you use ultra-sound), the wavelength is lambda = v / f = 2 cm. This means that your length measurement has an unambiguity range of 2 cm (sorry, cannot find a good link, google yourself). This means that you already need to know your distance to better than 2 cm, before you can use the result of your measurement to refine the distance.
Think of it in another way: when taking the cross-correlation between two perfect sines, the result should be a 'comb' of peaks with spacing equal to the wavelength. If they overlap perfectly, and you displace one signal by one wavelength, they still overlap perfectly. This means that you first have to know which of these peaks is the right one, otherwise a different peak can be the highest every time purely by random noise. Did you make a plot of the calculated cross-correlation before trying to blindly find the maximum?
This problem is the same as in interferometry, where it is easy to measure small distance variations with a resolution smaller than a wavelength by measuring phase differences, but you have no idea about the absolute distance, since you do not know the absolute phase.
The solution to this is actually easy: let your source generate more frequencies. Even using (band-limited) white-noise should work without problems when calculating cross-correlations, and it removes the ambiguity problem. You should see the white noise as a collection of sines. The cross-correlation of each of them will generate a comb, but with different spacing. When adding all those combs together, they will add up significantly only in a single point, at the delay you are looking for!
White Noise, Maximum Length Sequency or other non-periodic signals should be used as the test signal for time delay measurement using cross correleation. This is because non-periodic signals have only one cross correlation peak and there will be no ambiguity to determine the time delay. It is possible to use the burst type of periodic signals to do the job, but with degraded SNR. If you have to use a continuous periodic signal as the test signal, then you can only measure a time delay within one period of the periodic test signal. This should explain why, in your case, using lower frequency sine wave as the test signal works while using higher frequency sine wave does not. This is demonstrated in these videos: https://youtu.be/L6YJqhbsuFY, https://youtu.be/7u1nSD0RlwY .
I've been playing around a little with the Exocortex implementation of the FFT, but I'm having some problems.
Whenever I modify the amplitudes of the frequency bins before calling the iFFT the resulting signal contains some clicks and pops, especially when low frequencies are present in the signal (like drums or basses). However, this does not happen if I attenuate all the bins by the same factor.
Let me put an example of the output buffer of a 4-sample FFT:
// Bin 0 (DC)
FFTOut[0] = 0.0000610351563
FFTOut[1] = 0.0
// Bin 1
FFTOut[2] = 0.000331878662
FFTOut[3] = 0.000629425049
// Bin 2
FFTOut[4] = -0.0000381469727
FFTOut[5] = 0.0
// Bin 3, this is the first and only negative frequency bin.
FFTOut[6] = 0.000331878662
FFTOut[7] = -0.000629425049
The output is composed of pairs of floats, each representing the real and imaginay parts of a single bin. So, bin 0 (array indexes 0, 1) would represent the real and imaginary parts of the DC frequency. As you can see, bins 1 and 3 both have the same values, (except for the sign of the Im part), so I guess bin 3 is the first negative frequency, and finally indexes (4, 5) would be the last positive frequency bin.
Then to attenuate the frequency bin 1 this is what I do:
// Attenuate the 'positive' bin
FFTOut[2] *= 0.5;
FFTOut[3] *= 0.5;
// Attenuate its corresponding negative bin.
FFTOut[6] *= 0.5;
FFTOut[7] *= 0.5;
For the actual tests I'm using a 1024-length FFT and I always provide all the samples so no 0-padding is needed.
// Attenuate
var halfSize = fftWindowLength / 2;
float leftFreq = 0f;
float rightFreq = 22050f;
for( var c = 1; c < halfSize; c++ )
{
var freq = c * (44100d / halfSize);
// Calc. positive and negative frequency indexes.
var k = c * 2;
var nk = (fftWindowLength - c) * 2;
// This kind of attenuation corresponds to a high-pass filter.
// The attenuation at the transition band is linearly applied, could
// this be the cause of the distortion of low frequencies?
var attn = (freq < leftFreq) ?
0 :
(freq < rightFreq) ?
((freq - leftFreq) / (rightFreq - leftFreq)) :
1;
// Attenuate positive and negative bins.
mFFTOut[ k ] *= (float)attn;
mFFTOut[ k + 1 ] *= (float)attn;
mFFTOut[ nk ] *= (float)attn;
mFFTOut[ nk + 1 ] *= (float)attn;
}
Obviously I'm doing something wrong but can't figure out what.
I don't want to use the FFT output as a means to generate a set of FIR coefficients since I'm trying to implement a very basic dynamic equalizer.
What's the correct way to filter in the frequency domain? what I'm missing?
Also, is it really needed to attenuate negative frequencies as well? I've seen an FFT implementation where neg. frequency values are zeroed before synthesis.
Thanks in advance.
There are two issues: the way you use the FFT, and the particular filter.
Filtering is traditionally implemented as convolution in the time domain. You're right that multiplying the spectra of the input and filter signals is equivalent. However, when you use the Discrete Fourier Transform (DFT) (implemented with a Fast Fourier Transform algorithm for speed), you actually calculate a sampled version of the true spectrum. This has lots of implications, but the one most relevant to filtering is the implication that the time domain signal is periodic.
Here's an example. Consider a sinusoidal input signal x with 1.5 cycles in the period, and a simple low pass filter h. In Matlab/Octave syntax:
N = 1024;
n = (1:N)'-1; %'# define the time index
x = sin(2*pi*1.5*n/N); %# input with 1.5 cycles per 1024 points
h = hanning(129) .* sinc(0.25*(-64:1:64)'); %'# windowed sinc LPF, Fc = pi/4
h = [h./sum(h)]; %# normalize DC gain
y = ifft(fft(x) .* fft(h,N)); %# inverse FT of product of sampled spectra
y = real(y); %# due to numerical error, y has a tiny imaginary part
%# Depending on your FT/IFT implementation, might have to scale by N or 1/N here
plot(y);
And here's the graph:
The glitch at the beginning of the block is not what we expect at all. But if you consider fft(x), it makes sense. The Discrete Fourier Transform assumes the signal is periodic within the transform block. As far as the DFT knows, we asked for the transform of one period of this:
This leads to the first important consideration when filtering with DFTs: you are actually implementing circular convolution, not linear convolution. So the "glitch" in the first graph is not really a glitch when you consider the math. So then the question becomes: is there a way to work around the periodicity? The answer is yes: use overlap-save processing. Essentially, you calculate N-long products as above, but only keep N/2 points.
Nproc = 512;
xproc = zeros(2*Nproc,1); %# initialize temp buffer
idx = 1:Nproc; %# initialize half-buffer index
ycorrect = zeros(2*Nproc,1); %# initialize destination
for ctr = 1:(length(x)/Nproc) %# iterate over x 512 points at a time
xproc(1:Nproc) = xproc((Nproc+1):end); %# shift 2nd half of last iteration to 1st half of this iteration
xproc((Nproc+1):end) = x(idx); %# fill 2nd half of this iteration with new data
yproc = ifft(fft(xproc) .* fft(h,2*Nproc)); %# calculate new buffer
ycorrect(idx) = real(yproc((Nproc+1):end)); %# keep 2nd half of new buffer
idx = idx + Nproc; %# step half-buffer index
end
And here's the graph of ycorrect:
This picture makes sense - we expect a startup transient from the filter, then the result settles into the steady state sinusoidal response. Note that now x can be arbitrarily long. The limitation is Nproc > 2*min(length(x),length(h)).
Now onto the second issue: the particular filter. In your loop, you create a filter who's spectrum is essentially H = [0 (1:511)/512 1 (511:-1:1)/512]'; If you do hraw = real(ifft(H)); plot(hraw), you get:
It's hard to see, but there are a bunch of non-zero points at the far left edge of the graph, and then a bunch more at the far right edge. Using Octave's built-in freqz function to look at the frequency response we see (by doing freqz(hraw)):
The magnitude response has a lot of ripples from the high-pass envelope down to zero. Again, the periodicity inherent in the DFT is at work. As far as the DFT is concerned, hraw repeats over and over again. But if you take one period of hraw, as freqz does, its spectrum is quite different from the periodic version's.
So let's define a new signal: hrot = [hraw(513:end) ; hraw(1:512)]; We simply rotate the raw DFT output to make it continuous within the block. Now let's look at the frequency response using freqz(hrot):
Much better. The desired envelope is there, without all the ripples. Of course, the implementation isn't so simple now, you have to do a full complex multiply by fft(hrot) rather than just scaling each complex bin, but at least you'll get the right answer.
Note that for speed, you'd usually pre-calculate the DFT of the padded h, I left it alone in the loop to more easily compare with the original.
Your primary issue is that frequencies aren't well defined over short time intervals. This is particularly true for low frequencies, which is why you notice the problem most there.
Therefore, when you take really short segments out of the sound train, and then you filter these, the filtered segments wont filter in a way that produces a continuous waveform, and you hear the jumps between segments and this is what generates the clicks you here.
For example, taking some reasonable numbers: I start with a waveform at 27.5 Hz (A0 on a piano), digitized at 44100 Hz, it will look like this (where the red part is 1024 samples long):
So first we'll start with a low pass of 40Hz. So since the original frequency is less than 40Hz, a low-pass filter with a 40Hz cut-off shouldn't really have any effect, and we will get an output that almost exactly matches the input. Right? Wrong, wrong, wrong – and this is basically the core of your problem. The problem is that for the short sections the idea of 27.5 Hz isn't clearly defined, and can't be represented well in the DFT.
That 27.5 Hz isn't particularly meaningful in the short segment can be seen by looking at the DFT in the figure below. Note that although the longer segment's DFT (black dots) shows a peak at 27.5 Hz, the short one (red dots) doesn't.
Clearly, then filtering below 40Hz, will just capture the DC offset, and the result of the 40Hz low-pass filter is shown in green below.
The blue curve (taken with a 200 Hz cut-off) is starting to match up much better. But note that it's not the low frequencies that are making it match up well, but the inclusion of high frequencies. It's not until we include every frequency possible in the short segment, up to 22KHz that we finally get a good representation of the original sine wave.
The reason for all of this is that a small segment of a 27.5 Hz sine wave is not a 27.5 Hz sine wave, and it's DFT doesn't have much to do with 27.5 Hz.
Are you attenuating the value of the DC frequency sample to zero? It appears that you are not attenuating it at all in your example. Since you are implementing a high pass filter, you need to set the DC value to zero as well.
This would explain low frequency distortion. You would have a lot of ripple in the frequency response at low frequencies if that DC value is non-zero because of the large transition.
Here is an example in MATLAB/Octave to demonstrate what might be happening:
N = 32;
os = 4;
Fs = 1000;
X = [ones(1,4) linspace(1,0,8) zeros(1,3) 1 zeros(1,4) linspace(0,1,8) ones(1,4)];
x = ifftshift(ifft(X));
Xos = fft(x, N*os);
f1 = linspace(-Fs/2, Fs/2-Fs/N, N);
f2 = linspace(-Fs/2, Fs/2-Fs/(N*os), N*os);
hold off;
plot(f2, abs(Xos), '-o');
hold on;
grid on;
plot(f1, abs(X), '-ro');
hold off;
xlabel('Frequency (Hz)');
ylabel('Magnitude');
Notice that in my code, I am creating an example of the DC value being non-zero, followed by an abrupt change to zero, and then a ramp up. I then take the IFFT to transform into the time domain. Then I perform a zero-padded fft (which is done automatically by MATLAB when you pass in an fft size bigger than the input signal) on that time-domain signal. The zero-padding in the time-domain results in interpolation in the frequency domain. Using this, we can see how the filter will respond between filter samples.
One of the most important things to remember is that even though you are setting filter response values at given frequencies by attenuating the outputs of the DFT, this guarantees nothing for frequencies occurring between sample points. This means the more abrupt your changes, the more overshoot and oscillation between samples will occur.
Now to answer your question on how this filtering should be done. There are a number of ways, but one of the easiest to implement and understand is the window design method. The problem with your current design is that the transition width is huge. Most of the time, you will want as quick of transitions as possible, with as little ripple as possible.
In the next code, I will create an ideal filter and display the response:
N = 32;
os = 4;
Fs = 1000;
X = [ones(1,8) zeros(1,16) ones(1,8)];
x = ifftshift(ifft(X));
Xos = fft(x, N*os);
f1 = linspace(-Fs/2, Fs/2-Fs/N, N);
f2 = linspace(-Fs/2, Fs/2-Fs/(N*os), N*os);
hold off;
plot(f2, abs(Xos), '-o');
hold on;
grid on;
plot(f1, abs(X), '-ro');
hold off;
xlabel('Frequency (Hz)');
ylabel('Magnitude');
Notice that there is a lot of oscillation caused by the abrupt changes.
The FFT or Discrete Fourier Transform is a sampled version of the Fourier Transform. The Fourier Transform is applied to a signal over the continuous range -infinity to infinity while the DFT is applied over a finite number of samples. This in effect results in a square windowing (truncation) in the time domain when using the DFT since we are only dealing with a finite number of samples. Unfortunately, the DFT of a square wave is a sinc type function (sin(x)/x).
The problem with having sharp transitions in your filter (quick jump from 0 to 1 in one sample) is that this has a very long response in the time domain, which is being truncated by a square window. So to help minimize this problem, we can multiply the time-domain signal by a more gradual window. If we multiply a hanning window by adding the line:
x = x .* hanning(1,N).';
after taking the IFFT, we get this response:
So I would recommend trying to implement the window design method since it is fairly simple (there are better ways, but they get more complicated). Since you are implementing an equalizer, I assume you want to be able to change the attenuations on the fly, so I would suggest calculating and storing the filter in the frequency domain whenever there is a change in parameters, and then you can just apply it to each input audio buffer by taking the fft of the input buffer, multiplying by your frequency domain filter samples, and then performing the ifft to get back to the time domain. This will be a lot more efficient than all of the branching you are doing for each sample.
First, about the normalization: that is a known (non) issue. The DFT/IDFT would require a factor 1/sqrt(N) (apart from the standard cosine/sine factors) in each one (direct an inverse) to make them simmetric and truly invertible. Another possibility is to divide one of them (the direct or the inverse) by N, this is a matter of convenience and taste. Often the FFT routines do not perform this normalization, the user is expected to be aware of it and normalize as he prefers. See
Second: in a (say) 16 point DFT, what you call the bin 0 would correspond to the zero frequency (DC), bin 1 low freq... bin 4 medium freq, bin 8 to the highest frequency and bins 9...15 to the "negative frequencies". In you example, then, bin 1 is actually both the low frequency and medium frequency. Apart from this consideration, there is nothing conceptually wrong in your "equalization". I don't understand what you mean by "the signal gets distorted at low frequencies". How do you observe that ?