具有对数频率轴的 scipy 频谱图？

Question

Playing with scipy.signal.spectrogram. Works fine for what it is.

from scipy.io import wavfile
from scipy import signal
import numpy as np
import matplotlib.pyplot as plt

sf, audio = wavfile.read('serious.wav')
sig = np.mean(audio, axis=1)
f, t, Sxx = signal.spectrogram(sig, sf, scaling='spectrum')

plt.pcolormesh(t, f, np.log10(Sxx))
plt.ylabel('f [Hz]')
plt.xlabel('t [sec]')
plt.show()

This is the result:

But the frequency axis is linear. For audio this is not often desirable - at any rate, it's not what I want.

Is there a way to coax scipy.signal.spectrogram to output a logarithmic frequency scale?

If this is not doable with scipy, could you recommend an equally simple approach to obtain this result?

---

EDIT: The problem is not in the way the image is displayed. The problem is in the way the data is generated by signal.spectrogram()

I've changed the code like this:

plt.pcolormesh(t, f, np.log10(Sxx))
plt.ylabel('f [Hz]')
plt.xlabel('t [sec]')
plt.yscale('log')
plt.savefig('spec.png')
plt.show()

And now the image looks like this:

The f vector (generated by signal.spectrogram() ) looks like this:

array([    0.      ,   172.265625,   344.53125 ,   516.796875,
         689.0625  ,   861.328125,  1033.59375 ,  1205.859375,
        1378.125   ,  1550.390625,  1722.65625 ,  1894.921875,
        2067.1875  ,  2239.453125,  2411.71875 ,  2583.984375,
...
       19982.8125  , 20155.078125, 20327.34375 , 20499.609375,
       20671.875   , 20844.140625, 21016.40625 , 21188.671875,
       21360.9375  , 21533.203125, 21705.46875 , 21877.734375,
       22050.      ])

That's a linear distribution. I need far more points in the lower end, and far fewer at the top end.

Answer 1

I've found the problem. FFT is linear. My image is logarithmic. The default interval between frequencies is too big in the lower part of the frequency spectrum.

So I just upped the number of frequency samples via the nperseg parameter. In this example, the distance between successive frequencies is 1 Hz, which is pretty good resolution. Also, symlog scaling is best.

npts = int(sf)
f, t, Sxx = signal.spectrogram(sig, sf, nperseg=npts)
plt.yscale('symlog')

Of course, then there's too many frequencies at the top of the range, so some pruning is required within the f and Sxx arrays (dimensions must match, so prune them both the same way). Also, the range of displayed frequencies must be limited to 10 - 20000 or some reasonable values. All these optimizations are beyond the scope of this answer.

But I brought the script to the point where it's usable and I put it on GitHub:

https://github.com/FlorinAndrei/soundspec

Here's an example of a working spectrogram: