def test_error_checking(self):
self.assertRaises(ValueError,
lambda: segment_axis(np.arange(7), length=0,
shift=0))
self.assertRaises(ValueError,
lambda: segment_axis(np.arange(7), length=3,
shift=0))
def test_ending(self):
assert_equal(segment_axis(np.arange(6), length=3, shift=2, end='cut'),
np.array([[0, 1, 2], [2, 3, 4]]))
assert_equal(
segment_axis(
np.arange(6)+10, length=3, shift=2, end='pad', pad_mode='wrap'),
[[10, 11, 12], [12, 13, 14], [14, 15, 10]]
)
assert_equal(segment_axis(np.arange(6), length=3, shift=2, end='pad',
pad_value=-17),
np.array([[0, 1, 2], [2, 3, 4], [4, 5, -17]]))
def test_simple(self):
assert_equal(segment_axis(np.arange(6), length=3, shift=3),
np.array([[0, 1, 2], [3, 4, 5]]))
assert_equal(segment_axis(np.arange(7), length=3, shift=2),
np.array([[0, 1, 2], [2, 3, 4], [4, 5, 6]]))
assert_equal(segment_axis(np.arange(7), length=3, shift=1),
np.array([[0, 1, 2], [1, 2, 3], [2, 3, 4], [3, 4, 5],
[4, 5, 6]]))
assert_equal(segment_axis(np.arange(7), length=3, shift=-1),
[[4, 5, 6], [3, 4, 5], [2, 3, 4], [1, 2, 3], [0, 1, 2]])
def tbf_to_tbchw(x,
left_context,
right_context,
step_width,
pad_mode='symmetric',
pad_kwargs=None):
""" Transfroms data from TxBxF format to TxBxCxHxW format
This is only relevant for training a neural network in frames mode.
The abbreviations stand for:
T: Time frames
B: Batch size
F: Feature size
C: Channel (almost always 1)
H: Height of the convolution filter
W: Width of the convolution filter
:param x: Data to be transformed
:param left_context: Context size left to current frame
:param right_context: Context size right to current frame
:param step_width: Step width for window
:param pad_mode: Mode for padding. See :numpy.pad for details
:param pad_kwargs: Kwargs for pad call
:return: Transformed data
"""
if pad_kwargs is None:
pad_kwargs = dict()
x = np.pad(x, ((left_context, right_context), (0, 0), (0, 0)),
mode=pad_mode,
**pad_kwargs)
window_size = left_context + right_context + 1
return segment_axis(x, window_size, step_width, axis=0,
end='cut').transpose(0, 2, 3, 1)[:, :, None, :, :]
def est_time_shift(sig, ref_sig, seg_size, seg_shift):
"""Estimate the time shift between two signals
The time shift is estimated based on the generalized cross correlation with
phase transform (GCC-PhaT).
Args:
sig (numpy.ndarray):
Vector corresponding to a signal
ref_sig (numpy.ndarray):
Vector corresponding to the signal being used as reference
seg_size (int):
Size of the segments used in the GCC-PhaT algorithm
seg_shift:
Shift of the segments used in the GCC-PhaT algorithm
Returns:
shifts (numpy.ndarray):
Vector corresponding to the estimated time shifts
"""
def _get_gcpsd(seg, seg_ref):
"""Calculate the generalized cross power spectral density (GCPSD) for
the given signal segments
Args:
seg (array-like):
Vector corresponding to a segment of a signal
seg_ref (array-like):
Vector corresponding to the segment of the reference signal
Returns:
gcpsd (numpy.ndarray):
Vector corresponding to the GCPSD
"""
fft_seg = np.fft.fft(seg)
fft_ref_seg = np.fft.fft(seg_ref)
cpsd = np.conj(fft_ref_seg) * fft_seg
gcpsd = cpsd / (np.abs(fft_seg) * np.abs(fft_ref_seg) + 1e-18)
return gcpsd
segments = segment_axis(sig, seg_size, seg_shift, end='cut')
segments_ref = segment_axis(ref_sig, seg_size, seg_shift, end='cut')
shifts = np.zeros(len(segments))
for seg_idx, (seg, ref_seg) in enumerate(zip(segments, segments_ref)):
shifts[seg_idx] = max_time_lag_search(_get_gcpsd(seg, ref_seg))
return shifts
def test_multidimensional(self):
assert_equal(segment_axis(np.ones((2, 3, 4, 5, 6)), axis=3, length=3,
shift=2).shape,
(2, 3, 4, 2, 3, 6))
assert_equal(
segment_axis(np.ones((2, 3, 4, 5, 6)), axis=2, length=3, shift=2,
end='cut').shape,
(2, 3, 1, 3, 5, 6))
assert_equal(
segment_axis(np.ones((2, 3, 4, 5, 6)), axis=2, length=3, shift=2,
end='pad', pad_mode='wrap').shape,
(2, 3, 2, 3, 5, 6))
assert_equal(
segment_axis(np.ones((2, 3, 4, 5, 6)), axis=2, length=3, shift=2,
end='pad').shape,
(2, 3, 2, 3, 5, 6))
def maxfilt(x, n, axis=-1):
"""
Args:
x:
n:
axis:
Returns:
>>> x = np.ones((2, 5, 3)).cumsum(1)
>>> x[0] **= 2
>>> maxfilt(x, 3, axis=1).shape
"""
assert n % 2 == 1, n
if axis < 0:
axis = x.ndim + axis
pad_width = [[0, 0] for _ in range(x.ndim)]
pad_width[axis] = [(n - 1) // 2, (n - 1) // 2]
x = np.pad(x, pad_width, mode="constant")
x = segment_axis(x, n, shift=1, axis=axis, pad_mode="cut").max(axis + 1)
return x
def SRMR(signal: np.ndarray,
sample_rate: int = 16000,
n: int = 23,
low_freq: int = 125) -> float:
"""Python implementation of the SRMR metric.
Matlab reference implementation: https://github.com/MuSAELab/SRMRToolbox
Because results of other openly available SRMR python packages significantly deviate from
the original evaluation tool, this function reimplements the Matlab functionality.
An ASL-adjustment is not implemented, so that the calculated values still slightly differ from the Matlab implementation.
For an exact reproduction of the matlab results, the usage of an ASL-adjustion is required. However the deviation of
this implmentation from the Matlab version typically is not larger than 1e-3.
:param signal: signal on which the SRMR is calculated
:param sample_rate: sample rate of signal
:param n: number of gammatone filters used
:param low_freq: lowest center frequency of the gammatone filterbank, highest frequency is half the sample rate
:return: SRMR metric for given signal
"""
#Preprocessing of the signal (Voice activity detection)
signal = _preprocessing_vad(signal, sample_rate)
signal = signal - np.mean(signal)
signal /= np.std(signal, keepdims=True)
#Gammatone filterbank (with n Filters)
signal = gammatone_filterbank(signal,
sample_rate=sample_rate,
n=n,
low_freq=low_freq)
#Calculate temporal envelope of the signal
for i in range(len(signal)):
signal[i] = np.abs(sp.signal.hilbert(signal[i]))
#Frequencies of the modulation filters
modulation_filter_frequencies = [
4.0, 6.5, 10.7, 17.6, 28.9, 47.5, 78.1, 128.0
]
#Using 8 modulation filters on the output of the gammatone filters
E = []
for j in range(len(signal)):
E.append([])
for k in range(8):
W0 = math.tan(2 * math.pi * modulation_filter_frequencies[k] /
(2 * sample_rate))
B0 = W0 / 2
b = np.ndarray(
(3, ),
dtype=float,
buffer=np.array(
[B0 / (1 + B0 + W0**2), 0, -B0 / (1 + B0 + W0**2)]))
a = np.ndarray((3, ),
dtype=float,
buffer=np.array([
1, (2 * W0**2 - 2) / (1 + B0 + W0**2),
(1 - B0 + W0**2) / (1 + B0 + W0**2)
]))
E[j].append(sp.signal.lfilter(b, a, signal[j], axis=0))
#Calculation of the energy of the single bands
energy = []
for j in range(len(E)):
energy.append([])
for k in range(len(E[j])):
energy[j].append([])
#Segmentation of the signal
temp = segment_axis(E[j][k],
int(sample_rate / 1000) * 256,
int(sample_rate / 1000) * 64)
#Multiplication of a hamming window with each segment and summation of the result
hamm_window = sp.signal.hamming(int(sample_rate / 1000) * 256,
sym=True)
for window in temp:
energy[j][k].append(np.sum(np.square(hamm_window * window)))
#Calculation of the center frequencys (ERBS) and the corresponding ERBs
cfs = calculate_cfs(low_freq, sample_rate / 2, n)
ERBs = []
for i in range(len(cfs)):
ERBs.append(cfs[i] / 9.26449 + 24.7)
#Calculation of the means of the single bands
means = np.ndarray((len(energy), len(energy[0])))
for j in range(len(energy)):
for k in range(len(energy[j])):
means[j][k] = np.mean(energy[j][k])
#Calculation of the Bandwidth
total_energy = np.sum(np.sum(means))
AC_energy = np.sum(means, axis=1)
AC_perc = AC_energy * 100 / total_energy
sum = 0.0
BW = 0.0
for i in range(len(AC_perc)):
sum += AC_perc[i]
if (sum > 90):
BW = ERBs[i]
break
#Calculate cutoffs
cutoffs = []
for cfs in modulation_filter_frequencies:
w0 = 2 * math.pi * cfs / sample_rate
B0 = math.tan(w0 / 2) / 2
cutoffs.append(cfs - (B0 * sample_rate / (2 * math.pi)))
#Calculation of the mean of the different bands wth regards to the cuffoff band
numerator = np.sum(np.sum(means, axis=0)[:4])
denominator = np.sum(means, axis=0)[4]
for i in range(5, 8):
denominator += np.sum(means, axis=0)[i]
if cutoffs[i - 1] < BW < cutoffs[i]:
break
return numerator / denominator