def test_lifter_ceps(lifter_coefficient):
cepstra = np.linspace((1, 2), (10, 20), 10)
liftered_cepstra = lifter_ceps(cepstra, lifter_coefficient)
# check when liftering should be applied
if lifter_coefficient > 0:
n = np.arange(np.shape(cepstra)[1])
lift = 1 + (lifter_coefficient / 2.) * np.sin(
np.pi * n / lifter_coefficient)
np.testing.assert_almost_equal(liftered_cepstra, cepstra * lift, 3)
# check when liftering is not supposed to be applied
if lifter_coefficient == 0:
np.testing.assert_almost_equal(liftered_cepstra, cepstra, 3)
def test_lpcc(sig, fs, num_ceps, lifter, normalize):
"""
test LPCC features module for the following:
- check that the returned number of cepstrums is correct.
- check normalization.
- check liftering.
"""
lpccs = lpcc(sig=sig,
fs=fs,
num_ceps=num_ceps,
lifter=lifter,
normalize=normalize)
# assert number of returned cepstrum coefficients
if not lpccs.shape[1] == num_ceps:
raise AssertionError
# TO FIX: normalize
if normalize:
np.testing.assert_almost_equal(
lpccs,
cmvn(
cms(
lpcc(sig=sig,
fs=fs,
num_ceps=num_ceps,
lifter=lifter,
normalize=False))), 0)
else:
# TO FIX: lifter
if lifter > 0:
np.testing.assert_almost_equal(
lpccs,
lifter_ceps(
lpcc(sig=sig,
fs=fs,
num_ceps=num_ceps,
lifter=False,
normalize=normalize), lifter), 0)
def test_pncc(sig, fs, num_ceps, nfilts, nfft, low_freq, high_freq, dct_type,
use_energy, lifter, normalize):
"""
test PNCC features module for the following:
- check if ParameterErrors are raised for:
- nfilts < num_ceps
- negative low_freq value
- high_freq > fs / 2
- check that the returned number of cepstrums is correct.
- check the use energy functionality.
- check normalization.
- check liftering.
"""
# check error for number of filters is smaller than number of cepstrums
with pytest.raises(ParameterError):
pnccs = pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=num_ceps - 1,
nfft=nfft,
low_freq=low_freq,
high_freq=high_freq)
# check lifter Parameter error for low freq
with pytest.raises(ParameterError):
pnccs = pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=nfilts,
nfft=nfft,
low_freq=-5,
high_freq=high_freq)
# check lifter Parameter error for high freq
with pytest.raises(ParameterError):
pnccs = pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=nfilts,
nfft=nfft,
low_freq=low_freq,
high_freq=16000)
# compute features
pnccs = pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=nfilts,
nfft=nfft,
low_freq=low_freq,
high_freq=high_freq,
dct_type=dct_type,
use_energy=use_energy,
lifter=lifter,
normalize=normalize)
# assert number of returned cepstrum coefficients
if not pnccs.shape[1] == num_ceps:
raise AssertionError
# check use energy
if use_energy:
pnccs_energy = pnccs[:, 0]
gfccs_energy = pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=nfilts,
nfft=nfft,
low_freq=low_freq,
high_freq=high_freq,
dct_type=dct_type,
use_energy=use_energy,
lifter=lifter,
normalize=normalize)[:, 0]
np.testing.assert_almost_equal(pnccs_energy, gfccs_energy, 3)
# check normalize
if normalize:
np.testing.assert_almost_equal(
pnccs,
cmvn(
cms(
pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=nfilts,
nfft=nfft,
low_freq=low_freq,
high_freq=high_freq,
dct_type=dct_type,
use_energy=use_energy,
lifter=lifter,
normalize=False))), 3)
else:
# check lifter
if lifter > 0:
np.testing.assert_almost_equal(
pnccs,
lifter_ceps(
pncc(sig=sig,
fs=fs,
num_ceps=num_ceps,
nfilts=nfilts,
nfft=nfft,
low_freq=low_freq,
high_freq=high_freq,
dct_type=dct_type,
use_energy=use_energy,
lifter=False,
normalize=normalize), lifter), 3)
def lpcc(sig,
fs=16000,
num_ceps=13,
pre_emph=1,
pre_emph_coeff=0.97,
win_type="hann",
win_len=0.025,
win_hop=0.01,
do_rasta=True,
lifter=1,
normalize=1,
dither=1):
"""
Compute the LINEAR PREDICTIVE CEPSTRAL COEFFICIENTS (LPCC) from an audio signal.
Args:
sig (array) : a mono audio signal (Nx1) from which to compute features.
fs (int) : the sampling frequency of the signal we are working with.
Default is 16000.
num_ceps (float) : number of cepstra to return (order of the model to compute).
Default is 13.
pre_emph (int) : apply pre-emphasis if 1.
Default is 1.
pre_emph_coeff (float) : apply pre-emphasis filter [1 -pre_emph] (0 = none).
Default is 0.97.
win_type (float) : window type to apply for the windowing.
Default is hanning.
win_len (float) : window length in sec.
Default is 0.025.
win_hop (float) : step between successive windows in sec.
Default is 0.01.
do_rasta (int) : if 1 then apply rasta filtering.
Default is 0.
lifter (int) : apply liftering if value > 0.
Default is 22.
normalize (int) : apply normalization if 1.
Default is 0.
dither (int) : 1 = add offset to spectrum as if dither noise.
Default is 0.
Returns:
(array) : 2d array of LPCC features (num_frames x num_ceps)
"""
lpcs = lpc(sig=sig,
fs=fs,
num_ceps=num_ceps,
pre_emph=pre_emph,
pre_emph_coeff=pre_emph_coeff,
win_len=win_len,
win_hop=win_hop,
do_rasta=True,
dither=dither)
lpccs = lpc2cep(lpcs.T)
# liftering
if lifter > 0:
lpccs = lifter_ceps(lpccs, lifter)
# normalization
if normalize:
lpccs = cmvn(cms(lpccs))
lpccs = lpccs.T
return lpccs[:, :]
