Python adaptfilt 2.0 FloatingPointError: invalid value encountered in multiply - python-3.x

Question

I am trying to use the "echo cancel" example in python 3.4 from the library adaptfilt 2.0 which looks like this:
import numpy as np
import adaptfilt as adf
# Get u(n) - this is available on github or pypi in the examples folder
u = np.load('speech.npy')
# Generate received signal d(n) using randomly chosen coefficients
coeffs = np.concatenate(([0.8], np.zeros(8), [-0.7], np.zeros(9),
[0.5], np.zeros(11), [-0.3], np.zeros(3),
[0.1], np.zeros(20), [-0.05]))
d = np.convolve(u, coeffs)
# Add background noise
v = np.random.randn(len(d)) * np.sqrt(5000)
d += v
# Apply adaptive filter
M = 100 # Number of filter taps in adaptive filter
step = 0.1 # Step size
y, e, w = adf.nlms(u, d, M, step, returnCoeffs=True)
# Calculate mean square weight error
mswe = adf.mswe(w, coeffs)
It works as expected. But then I wanted to to do the same thing with some real data from music file and I get an error:
Traceback (most recent call last):
File "C:/Python34/Lib/site-packages/adaptfilt/echocancel.py", line 86, in <module>
y, e, w = adf.nlms(u, d, M, step, returnCoeffs=True)
File "C:\Python34\Lib\site-packages\adaptfilt\nlms.py", line 149, in nlms
w = leakstep * w + step * normFactor * x * e[n]
FloatingPointError: invalid value encountered in multiply
The code I used is this:
import numpy as np
import adaptfilt as adf
import pyaudio
import wave
np.seterr(all='raise')
p = pyaudio.PyAudio()
stream = p.open(format = p.get_format_from_width(2),
channels = 1,
rate = 44100,
input = True,
output = True,
#stream_callback = self.callback
)
wf = wave.open("XXX.wav", 'rb')
while u != " ":
data = wf.readframes(1024)
u = np.fromstring(data, np.int16)
# Generate received signal d(n) using randomly chosen coefficients
coeffs = np.concatenate(([0.8], np.zeros(8), [-0.7], np.zeros(9),
[0.5], np.zeros(11), [-0.3], np.zeros(3),
[0.1], np.zeros(20), [-0.05]))
coeffs.dtype = np.int16
d = np.convolve(u, coeffs)
# Add background noise
v = np.random.randn(len(d)) * np.sqrt(5000)
d += v
# Apply adaptive filter
M = 100 # Number of filter taps in adaptive filter
step = 0.1 # Step size
y, e, w = adf.nlms(u, d, M, step, returnCoeffs=True)
# Calculate mean square weight error
mswe = adf.mswe(w, coeffs)
stream.write(y.astype(np.int16).tostring())
The only difference I see is that the array from "speech.npy" is type of float64 and my array from the wav file is type of int16.

Answer 1

I was also able to get 'adaptfilt 2.0' to work on music data (a mono .wav file with 44.1kHz sampling rate) by casting 'data' to float64.
The adaptive filter took longer to converge, but it worked OK. Below is the Mean squared weight error plot showing the filter converge.
I would also add that to apply this to music, you'll likely need a longer filter tap (M). The 'speech.npy' array used in the original script is just an array with no sample rate information, but we can assume the sample rate in a speech file is lower than 44.1kHz.
I played the 'speech.npy' back at varying sample rates, and just by listening to what sounds natural, I'm guessing it's in the 10-12kHz range. This means that the "impulse response" stored in 'coeffs', is ~5.7 ms (assuming 10kHz sampling rate). At 44.1kHz, the impulse response is only ~1.3ms. This is very short and unlikely to model the impulse response for real signals.

Answer 2

This error arises due to numerical issues when using the int16 type for computations somewhere inside the module. Using your example and casting the input data to a floating point type such as
u = np.fromstring(data, np.int16)
u = np.float64(u)
resolves the issue for me.