def __init__(self,
sample_rate: int = 10000,
use_vad: bool = True,
extended: bool = False):
super().__init__()
# Independant from FS
self.sample_rate = sample_rate
self.use_vad = use_vad
self.extended = extended
self.intel_frames = N
self.beta = BETA
self.dyn_range = DYN_RANGE
self.do_resample = True
# Dependant from FS
if self.do_resample:
sample_rate = FS
self.resample = torchaudio.transforms.Resample(
orig_freq=self.sample_rate,
new_freq=FS,
resampling_method='sinc_interpolation',
)
self.win_len = (N_FRAME * sample_rate) // FS
self.nfft = 2 * self.win_len
win = torch.from_numpy(np.hanning(self.win_len + 2)[1:-1]).float()
self.win = nn.Parameter(win, requires_grad=False)
obm_mat = thirdoct(sample_rate, self.nfft, NUMBAND, MINFREQ)[0]
self.OBM = nn.Parameter(torch.from_numpy(obm_mat).float(),
requires_grad=False)
def test_thirdoct():
"""Test thirdoct by comparing to Octave"""
obm_m, cf_m = octave.feval('octave/thirdoct.m',
float(FS), float(NFFT), float(NUMBAND),
float(MINFREQ), nout=2)
obm, cf = thirdoct(FS, NFFT, NUMBAND, MINFREQ)
obm_m = np.array(obm_m)
cf_m = np.array(cf_m).transpose().squeeze()
assert_allclose(obm, obm_m, atol=ATOL)
assert_allclose(cf, cf_m, atol=ATOL)
def test_thirdoct():
obm_m, cf_m = eng.thirdoct(float(FS),
float(NFFT),
float(NUMBAND),
float(MINFREQ),
nargout=2)
obm, cf = thirdoct(FS, NFFT, NUMBAND, MINFREQ)
obm_m = np.array(obm_m)
cf_m = np.array(cf_m).transpose().squeeze()
assert_allclose(obm, obm_m, atol=ATOL)
assert_allclose(cf, cf_m, atol=ATOL)