# Gain error when performing STFT spectogram of increasing frequency signal

I'm performing a STFT of a sine wave in which the frequency increases (up-chirp signal) using python numpy.fft as shown in the following code:

import numpy as np
import matplotlib.pyplot as plt

def stft(x, L, overlap, window):
opoints = int(L*overlap)
X = np.array([np.abs(np.fft.fft(window*x[i:i+L])[:L//2])
for i in range(0,len(x)-L,opoints)])
return X

t = np.linspace(0, 5, 11025*5)

xf = np.linspace(0, 2000, 11025*5)
w = 2*np.pi*xf
x = np.sin(w*t)

f = np.linspace(0, 11025, 1024)

window = np.hanning(1024)

stft = stft(x,1024,0.5,window)

plt.imshow(stft[:,:1024//2].T, cmap = plt.get_cmap('gist_yarg'), origin='lower', aspect='auto',
interpolation='nearest', extent=(0,5, 0,11025/2))

plt.xlabel("Time [s]")
plt.ylabel("Frequency [Hz]")
plt.show()


It works for a constant frequency sinewave. However it shows not an offset but a gain error if using the up-chirp signal as input.

STFT of increasing frequency sine wave

I'm using:

• Hanning window
• Overlap = 0.5
• Window size 1024
• Sampling frequency 11025Hz
• Total sample time: 5 seconds

What is causing the gain error that leads the 2 kHz signal appearing at double frequency at the end of the time window?

A sinusoidal signal is

$$y(t) = sin(\phi(t))$$

The phase of a linear chirp, is

$$\phi(t) = \phi_0 + 2\pi\left(f_0t+\frac{k}{2}t^2\right)$$ where $$k = \frac{f_1-f_0}{t_1-t_0}$$

So your chirp signal is missing a 0.5 correction for the frequency variation.

Use the scipy.signal.chirp:

import numpy as np
import matplotlib.pyplot as plt
from scipy.signal import chirp

def stft(x, L, overlap, window):
opoints = int(L*overlap)
X = np.array([np.abs(np.fft.fft(window*x[i:i+L])[:L//2])
for i in range(0,len(x)-L,opoints)])
return X

tmax = 5
fs = 11025
N = fs*tmax

t = np.linspace(0, tmax, N)
x = chirp(t,0,tmax,2000)
windowsize = 1024

f = np.linspace(0, fs, windowsize)
window = np.hanning(windowsize)
stft = stft(x,windowsize,0.5,window)

plt.imshow(stft.T, cmap = plt.get_cmap('gist_yarg'), origin='lower', aspect='auto',
interpolation='nearest', extent=(0,t[-1], 0,fs/2))
plt.xlabel("Time [s]")
plt.ylabel("Frequency [Hz]")
plt.show() Sources:

• You are right, the bug was causing the gain error not to show when using overlap < 0.05 since all the signal data was not being displayed. I edited the first message. Thank you for the contribution, still trying to figure out what is causing the gain error. Oct 24, 2017 at 14:18
• Thank you so much @daguiam, again you are right, I was not understanding the chirp signal properly. Oct 24, 2017 at 22:55