def test_preemphasis():
for dtype in [np.float32, np.float64]:
np.random.seed(1234)
x = np.random.rand(16000 * 5).astype(dtype)
y = preemphasis(x, 0.97)
assert x.shape == y.shape
assert x.dtype == y.dtype
x_hat = inv_preemphasis(y, 0.97)
assert x_hat.dtype == x.dtype
assert np.allclose(x_hat, x, atol=1e-5)
x = np.random.rand(16000)
assert np.allclose(P.preemphasis(x, 0), x)
assert np.allclose(P.inv_preemphasis(P.preemphasis(x, 0), 0), x)
def preemphasis(x):
# allow disable preemphasis
if hparams.preemphasis == -1.0:
return x
else:
from nnmnkwii.preprocessing import preemphasis
return preemphasis(x, hparams.preemphasis)
def melspectrogram(y, ref_level_db, allow_clipping_in_normalization,
min_level_db):
D = _lws_processor().stft(preemphasis(y)).T
S = _amp_to_db(_linear_to_mel(np.abs(D))) - ref_level_db
if not allow_clipping_in_normalization:
assert S.max() <= 0 and S.min() - min_level_db >= 0
return _normalize(S)
def wave2spec(
wave, fs, frame_period, window,
nperseg=None, nmels=80, preemphasis_coef=None,
f_min=0, f_max=None, dtype='float32',
return_t=False):
stft_kwargs = make_stft_args(
frame_period, fs, nperseg=nperseg, window=window)
htk, norm = True, "slaney"
if preemphasis_coef is not None:
spec_wave = preemphasis(wave, preemphasis_coef)
else:
spec_wave = wave
_, t, Zxx = signal.stft(
spec_wave, **stft_kwargs)
pspec = np.abs(Zxx)
mspec = melspectrogram(
sr=fs, S=pspec,
n_mels=nmels, fmin=f_min, fmax=f_max,
power=1.0, htk=htk, norm=norm)
pspec = pspec.T.astype(dtype)
mspec = mspec.T.astype(dtype)
upsample = fs // (1000 // frame_period)
length = (len(wave) // upsample) * upsample
wave = wave[:length]
mspec = mspec[:length // upsample]
spec = pspec[:length // upsample]
if return_t:
return wave, spec, mspec, upsample, t
else:
return wave, spec, mspec, upsample
def spectrogram(y):
D = librosa.stft(preemphasis(y),
n_fft=hparams.fft_size,
hop_length=hparams.hop_size,
win_length=hparams.fft_wsize)
S = _amp_to_db(np.abs(D)) - hparams.spec_ref_level_db
#S = librosa.amplitude_to_db(np.abs(D)) - hparams.spec_ref_level_db
return _normalize(S)
def melspectrogram(y):
D = librosa.stft(preemphasis(y),
n_fft=hparams.fft_size,
hop_length=hparams.hop_size,
win_length=hparams.fft_wsize)
S = _amp_to_db(_linear_to_mel(np.abs(D))) - hparams.spec_ref_level_db
#S = librosa.amplitude_to_db(_linear_to_mel(np.abs(D))) - hparams.spec_ref_level_db
if not hparams.allow_clipping_in_normalization:
assert S.max() <= 0 and S.min() - hparams.min_level_db >= 0
return _normalize(S)
def melspectrogram(y):
a = preemphasis(y)
# stft fails if len(a.shape) != 1, not sure how it got there
if len(a.shape) != 1:
a = np.squeeze(a)
D = _lws_processor().stft(a).T
S = _amp_to_db(_linear_to_mel(np.abs(D))) - hparams.ref_level_db
if not hparams.allow_clipping_in_normalization:
assert S.max() <= 0 and S.min() - hparams.min_level_db >= 0
return _normalize(S)
def spectrogram(y, power, pcen=False):
global _mel_freqs
stftS = librosa.stft(y,
n_fft=hparams.fft_size,
hop_length=hparams.hop_size)
if hparams.use_preemphasis:
y = preemphasis(y)
S = librosa.stft(y, n_fft=hparams.fft_size, hop_length=hparams.hop_size)
if _mel_freqs is None:
_mel_freqs = librosa.mel_frequencies(S.shape[0], fmin=hparams.fmin)
_S = librosa.perceptual_weighting(np.abs(S)**power,
_mel_freqs,
ref=hparams.ref_level_db)
return _normalize(_S - hparams.ref_level_db), stftS
def melspectrogram(y):
D = _stft(preemphasis(y))
S = _amp_to_db(_linear_to_mel(np.abs(D)**hparams.magnitude_power)) - hparams.ref_level_db
if not hparams.allow_clipping_in_normalization:
assert S.max() <= 0 and S.min() - hparams.min_level_db >= 0
return _normalize(S)
def melspectrogram(y):
D = _stft(preemphasis(y))
S = _amp_to_db(_linear_to_mel(np.abs(D))) - hparams.ref_level_db
return _normalize(S)
def preemphasis(x, coef=0.85):
return P.preemphasis(x, coef)
def preemphasis(x):
from nnmnkwii.preprocessing import preemphasis
return preemphasis(x, preemphasis_factor)
def melspectrogram(y):
D = _lws_processor().stft(preemphasis(y)).T
S = _amp_to_db(_linear_to_mel(np.abs(D)))
return _normalize(S)
def stft(y, preemp=False):
if preemp:
y = preemphasis(y)
return librosa.stft(y, n_fft=hparams.fft_size, hop_length=hparams.hop_size)
def preemphasis(x):
from nnmnkwii.preprocessing import preemphasis
return preemphasis(x, hparams.preemphasis)
def spectrogram(y):
D = _lws_processor().stft(preemphasis(y)).T
S = _amp_to_db(np.abs(D)) - hparams.ref_level_db
return _normalize(S)
def _preemphasis(self, x):
from nnmnkwii.preprocessing import preemphasis
return preemphasis(x, self.hparams.preemphasis)
