Best way to filter out frequencies in real time (frame by frame) analysis ?

How can I filter out particular frequencies in real time frame by frame analysis of music (sound) data ?

tried a butter filter, but it doesn't transition properly between the frames (I do hear ticks). As I understand fft and ifft is not the best way to go, also, since it fails to properly transition between blocks ?

What is the best way to deal with this ?

A real code example is equal to gold to me, since I am a noob when dealing with MatLab.

[magate]

Have you tried designing a filter? Are you trying to do a low pass filter?

[Jan]

Please post what you have tried with the Butterworth filter. It does not matter if you filter the signal in blocks or in total, as long as you catch the 2nd output if filter and provide it as input for the next block.

[Remigijus Rebliavicius]

Jan Simon thanks ! I didn't know I need to pass the previous conditions to the next frame filtered... Now it works as it should !

[Star Strider]

For real time applications, you need the DSP System Toolbox (link) or the Audio System Toolbox (link).

Answer 1

[Jan]

And for clarity a code example (UNTESTED CODE):

signal    = rand(1, 1000);
[b, a]    = butter(3, 0.5);
z         = zeros(1, 3);
filtered1 = filter(b, a, signal, z);  % Same as: filter(b, a, signal) if z is 0

filtered2 = NaN(size(signal));
steps     = 0:100:numel(signal);
for iStep = 1:numel(steps) - 1
  i1 = steps(iStep) + 1;
  i2 = steps(iStep + 1);
  % z is updated in each block:
  [filtered2(i1:i2), z] = filter(b, a, signal(i1:i2), z);
end
max(abs(filtered1 - filtered2))

This shows, that using the final state of the filter z as initial state for the next input is equivalent to filtering the complete signal at once.

[Artem Bobritsky]

Very nice solution and it works as intended!

One problem I have is:

If you design the filter using the [b,a] syntax, you might encounter numerical problems. These problems are due to round-off errors and can occur for n as low as 4.

And every filter above 4th order corrupts the signal.

Is there any chance to use similar approach with [z,p,k] or digitalFilter?

Thanks

[Jan]

@Artem: The less the readers have to guess, the more likely my answer will match your needs. Please post, which filter command you want to split. I'm not experienced in DSP.

[Artem Bobritsky]

@Jan,

Thank you for such prompt reply.

I'm lost in Matlab world completely....

I'm trying to make

[bbp, dfBandpas]  = bandpass(signal,[0.2 4], fs,'ImpulseResponse','iir');
filtered1 = filter(dfBandpas, signal); 

as realtime.

In my case, it generates eliptic filter of 14th order. I can easily convert the output to the transfer function [b,a] and use your approach.

But the problem if order is higher than 4 - [b,a] don't have anough precition... And I have severe distortions in the filtered signal.

[Jan]

@Artem: Do the DSP System Toolbox (link) or the Audio System Toolbox (link) offer, what you need for real time filtering?

[Artem Bobritsky]

@Jan,

Thank you for answering this!

I went through DSP System Toolbox and found Butterworth example for Simulink and nothing for Matlab.. I'll take a look with more attention.

I'm filtering EEG signals (Delta, Alpha, etc.) using band pass filters. Post processing works quite good, but it's time to move it into a microcontroller, so I have no way to collect 8 hours of data and post process it, I need to make some decitions based on the filtered signal.

I thought I could use a code gen for that... Unfortunately it's so limited and full of hardcoded constants, e.g. butter(_) needs _constants_ !!! bandpass(_) could not be used for code generation.