Equivalent matlab function for python librosa.resample

I need to upsample my original data of 250Hz to 96000Hz.I did it in python using librosa.resample. But when I did it in matlab I am not getting a smooth signal.I need to get an exact replica of upsampled output that I got in python.

Here i am attaching a data.mat file. In the data.mat, UL_filter is the input at sampling rate 250Hz and x is the upsampled signal which i got in python. I need to upsample the UL_filter signal @250Hz to 96000Hz and which should be same to my python output. In python, i have used the code librosa.resample(UL_filter,250,96000) and in matlab i have used resample.

To get an exact replica of python upsampled output , what function I have to use in matlab? Is there any equivalent function available in matlab for librosa.resample?

Answer 1

[Mathieu NOE]

hello Betty

you can do this to upsample the data :

n = length(UL_filter);
nx = length(x);

% code for upsampling data 
x_original = (0:n-1);
fs_original = 250;
fs_new = 96000; 
fs_ratio = fs_new/fs_original;
x_new = linspace(0,n-1,n*fs_ratio);

upsampled_data = interp1(x_original,UL_filter,x_new,'linear');

plot(x_original,UL_filter,'*b',x_new,upsampled_data)
legend('original data','upsampled data');

[Betty Kurian]

Thank you for the answer. I tried this code, the ruggedness of the curve reduced. But I need an upsampled output which is exactly similar to the python output(x in the data.mat). In librosa.resample,by default, this uses a high-quality (but relatively slow) method (‘kaiser_best’) for band-limited sinc interpolation. How can we do it in the MATLAB?

[Mathieu NOE]

hello

well, how close can we get to the python output, I don't know as I don't have the values python will generate

If you have those data , I can see where we are and if we need to refine the matlab method

are you able to provide a test case with the data python generates ?

[Betty Kurian]

hello,

Here I am attaching a figure which is the comparison plot of python and matlab upsampled outputs. I am getting a smooth curve (green) while upsampling in python whereas an oscillating output(red) in matlab. In data.mat, python upsampled output is given as 'x' and 'UL_filter' is the input. We need to upsample the UL_filter from 250Hz to 96000Hz and need to obtain an output similar to 'x'. I am trying to replicate the python output in matlab. Here I am attaching the documentation of librosa.resample that i have used in python to upsample.

https://librosa.org/doc/main/generated/librosa.resample.html

[Mathieu NOE]

hello

funny I don't see those "waves" in my linear interpolated data (red curve).

FYI, a linear interpolation will give "straight" lines between the 250 Hz sampled data - and this is what I get when I zoom in the same x axis interval as in your figure

Second, I modified a bit the code and selected also a spline interpolation method for a smoother ouput

code :

clc
clearvars
load('data.mat');

n = length(UL_filter);
nx = length(x);
% code for upsampling data 
fs_original = 250;
dt_original = 1/fs_original;

fs_new = 96000; 
dt_new = 1/fs_new;

time_axis_original = (0:n-1)*dt_original;
time_axis_new = (0:nx-1)*dt_new;

upsampled_data = interp1(time_axis_original,UL_filter,time_axis_new,'spline');  % select interpolation method among the list below
%     'linear'   - (default) linear interpolation
%     'spline'   - piecewise cubic spline interpolation (SPLINE)
%     'pchip'    - shape-preserving piecewise cubic interpolation
%     'cubic'    - cubic convolution interpolation for uniformly-spaced
%                  data. This method does not extrapolate and falls back to
%                  'spline' interpolation for irregularly-spaced data.
%            NOTE: 'cubic' changed in R2020b to perform cubic convolution.
%                  In previous releases, 'cubic' was the same as 'pchip'.
%     'v5cubic'  - same as 'cubic'
%     'makima'   - modified Akima cubic interpolation

plot(time_axis_original,UL_filter,'*b',time_axis_new,x,'*g',time_axis_new,upsampled_data,'r')
legend('original data','python data','matlab upsampled data');

plot :

now you can see that the 3 data sets do perfectly overlay - python and matlab outputs are now equivalent.

there are other options you can try as well as "cubic" or "makina". We can compute the error between python and matlab outputs , but I also noticed that python generates some side effects due to it's sinc window : did you see the oscillations at the beginning and the end of you resampled data ?

[Mathieu NOE]

hello Betty

if my answer has fullfilled your expectations, do you mind accepting it ?

thanks