def __test(scale, n_fmt, over_sample, kind, y_orig, y_res, atol):
# Make sure our signals preserve energy
assert np.allclose(np.sum(y_orig**2), np.sum(y_res**2))
# Scale-transform the original
f_orig = librosa.fmt(y_orig,
t_min=0.5,
n_fmt=n_fmt,
over_sample=over_sample,
kind=kind)
# Force to the same length
n_fmt_res = 2 * len(f_orig) - 2
# Scale-transform the new signal to match
f_res = librosa.fmt(y_res,
t_min=scale * 0.5,
n_fmt=n_fmt_res,
over_sample=over_sample,
kind=kind)
# Due to sampling alignment, we'll get some phase deviation here
# The shape of the spectrum should be approximately preserved though.
assert np.allclose(np.abs(f_orig), np.abs(f_res), atol=atol, rtol=1e-7)
def __test(scale, n_fmt, over_sample, kind, y_orig, y_res, atol):
# Make sure our signals preserve energy
assert np.allclose(np.sum(y_orig**2), np.sum(y_res**2))
# Scale-transform the original
f_orig = librosa.fmt(y_orig,
t_min=0.5,
n_fmt=n_fmt,
over_sample=over_sample,
kind=kind)
# Force to the same length
n_fmt_res = 2 * len(f_orig) - 2
# Scale-transform the new signal to match
f_res = librosa.fmt(y_res,
t_min=scale * 0.5,
n_fmt=n_fmt_res,
over_sample=over_sample,
kind=kind)
# Due to sampling alignment, we'll get some phase deviation here
# The shape of the spectrum should be approximately preserved though.
assert np.allclose(np.abs(f_orig), np.abs(f_res), atol=atol, rtol=1e-7)
def test_fmt_axis():
y = np.random.randn(32, 32)
f1 = librosa.fmt(y, axis=-1)
f2 = librosa.fmt(y.T, axis=0).T
assert np.allclose(f1, f2)
def test_fmt_axis():
y = np.random.randn(32, 32)
f1 = librosa.fmt(y, axis=-1)
f2 = librosa.fmt(y.T, axis=0).T
assert np.allclose(f1, f2)
def __call__(self):
self.on_launch()
xdata = []
ydata = []
for x in np.arange(0, 10, 0.5):
ydata = np.array([
np.exp(-i**2) + 10 * np.exp(-(i - 7)**2)
for i in range(0, 128)
])
if (x % 2 == 0):
ydata = np.abs(librosa.fmt(ydata, n_fmt=64))
else:
ydata = librosa.amplitude_to_db(librosa.stft(ydata),
ref=np.max)
xdata = np.array([i for i in range(0, ydata.size)])
self.on_running(xdata, ydata)
time.sleep(1)
return xdata, ydata
def transform_audio(self, y):
'''Apply the scale transform to the tempogram
Parameters
----------
y : np.ndarray
The audio buffer
Returns
-------
data : dict
data['temposcale'] : np.ndarray, shape=(n_frames, n_fmt)
The scale transform magnitude coefficients
'''
data = super(TempoScale, self).transform_audio(y)
data['temposcale'] = np.abs(
fmt(data.pop('tempogram'), axis=1,
n_fmt=self.n_fmt)).astype(np.float32)[self.idx]
return data
def compute_stm(counts, nbins=300, ncycles=5):
start_ind = 0
nsamples = nbins*ncycles
end_ind = start_ind + nsamples
n_acf = 1500
n_stm = 750 #int(nsamples/4.)
ac_all, scale_all = [], []
while end_ind <= len(counts):
ac = librosa.autocorrelate(counts[start_ind:end_ind], max_size=n_acf)
ac = librosa.util.normalize(ac, norm=np.inf)
ac_all.append(ac)
scale = librosa.fmt(ac, n_fmt=n_stm)
scale_all.append(scale)
start_ind += nsamples
end_ind += nsamples
ac_all = np.array(ac_all)
scale_all = np.array(scale_all)
return ac_all, scale_all
def __test(t_min, n_fmt, over_sample, y):
librosa.fmt(y, t_min=t_min, n_fmt=n_fmt, over_sample=over_sample)
def __test(t_min, n_fmt, over_sample, y):
librosa.fmt(y, t_min=t_min, n_fmt=n_fmt, over_sample=over_sample)
