def merlin_post_filter(mgc,
alpha,
minimum_phase_order=511,
fftlen=2048,
coef=1.4,
weight=None):
_, D = mgc.shape
if weight is None:
weight = np.ones(D) * coef
weight[:2] = 1
assert len(weight) == D
mgc_r0 = pysptk.c2acr(pysptk.freqt(mgc, minimum_phase_order, alpha=-alpha),
0, fftlen).flatten()
mgc_p_r0 = pysptk.c2acr(
pysptk.freqt(mgc * weight, minimum_phase_order, -alpha), 0,
fftlen).flatten()
mgc_b0 = pysptk.mc2b(weight * mgc, alpha)[:, 0]
mgc_p_b0 = np.log(mgc_r0 / mgc_p_r0) / 2 + mgc_b0
mgc_p_mgc = pysptk.b2mc(
np.hstack((mgc_p_b0[:, None], pysptk.mc2b(mgc * weight,
alpha)[:, 1:])), alpha)
return mgc_p_mgc
def test_merlin_post_filter():
root = join(DATA_DIR, "merlin_post_filter")
mgc = np.fromfile(join(root, "arctic_b0539.mgc"),
dtype=np.float32).reshape(-1, 60)
weight = np.fromfile(join(root, "weight"), dtype=np.float32)
alpha = 0.58
minimum_phase_order = 511
fftlen = 1024
coef = 1.4
# Step 1
mgc_r0 = np.fromfile(join(root, "arctic_b0539.mgc_r0"), dtype=np.float32)
mgc_r0_hat = pysptk.c2acr(pysptk.freqt(
mgc, minimum_phase_order, alpha=-alpha), 0, fftlen).flatten()
assert np.allclose(mgc_r0, mgc_r0_hat)
# Step 2
mgc_p_r0 = np.fromfile(
join(root, "arctic_b0539.mgc_p_r0"), dtype=np.float32)
mgc_p_r0_hat = pysptk.c2acr(pysptk.freqt(
mgc * weight, minimum_phase_order, -alpha), 0, fftlen).flatten()
assert np.allclose(mgc_p_r0, mgc_p_r0_hat)
# Step 3
mgc_b0 = np.fromfile(join(root, "arctic_b0539.mgc_b0"), dtype=np.float32)
mgc_b0_hat = pysptk.mc2b(weight * mgc, alpha)[:, 0]
assert np.allclose(mgc_b0, mgc_b0_hat)
# Step 4
mgc_p_b0 = np.fromfile(
join(root, "arctic_b0539.mgc_p_b0"), dtype=np.float32)
mgc_p_b0_hat = np.log(mgc_r0_hat / mgc_p_r0_hat) / 2 + mgc_b0_hat
assert np.allclose(mgc_p_b0, mgc_p_b0_hat)
# Final step
mgc_p_mgc = np.fromfile(
join(root, "arctic_b0539.mgc_p_mgc"), dtype=np.float32).reshape(-1, 60)
mgc_p_mgc_hat = pysptk.b2mc(
np.hstack((mgc_p_b0_hat[:, None], pysptk.mc2b(mgc * weight, alpha)[:, 1:])), alpha)
assert np.allclose(mgc_p_mgc, mgc_p_mgc_hat)
filtered_mgc = merlin_post_filter(mgc, alpha, coef=coef, weight=weight,
minimum_phase_order=minimum_phase_order,
fftlen=fftlen)
assert np.allclose(filtered_mgc, mgc_p_mgc, atol=1e-6)
def __test_synthesis_levdur(filt):
# dummy source excitation
source = __dummy_source()
hopsize = 80
# dummy filter coef.
windowed = __dummy_windowed_frames(source,
frame_len=512,
hopsize=hopsize)
c = pysptk.mcep(windowed, filt.order)
lpc = pysptk.levdur(pysptk.c2acr(c))
# make sure lpc has loggain
lpc[:, 0] = np.log(lpc[:, 0])
# synthesis
synthesizer = Synthesizer(filt, hopsize)
y = synthesizer.synthesis(source, lpc)
assert np.all(np.isfinite(y))
def pitch_shift_on_lpc_residual(
wav,
sr,
shift_in_cent,
frame_length=4096,
hop_length=240,
mgc_order=59,
):
assert wav.dtype == np.int16
frames = (librosa.util.frame(wav,
frame_length=frame_length,
hop_length=hop_length).astype(np.float64).T)
frames *= pysptk.blackman(frame_length)
alpha = pysptk.util.mcepalpha(sr)
mgc = pysptk.mcep(frames, mgc_order, alpha, eps=1e-5, etype=1)
c = pysptk.freqt(mgc, mgc_order, -alpha)
lpc = pysptk.levdur(pysptk.c2acr(c, mgc_order, frame_length))
# remove gain
lpc[:, 0] = 0
# Compute LPC residual
synth = Synthesizer(AllZeroDF(mgc_order), hop_length)
wav_lpc = synth.synthesis(wav.astype(np.float64), -lpc)
residual = wav - wav_lpc
# Pitch-shift on LPC residual
residual_shifted = librosa.effects.pitch_shift(residual,
sr=sr,
n_steps=shift_in_cent,
bins_per_octave=1200)
# Filtering by LPC
synth = Synthesizer(AllPoleDF(mgc_order), hop_length)
wav_shifted = synth.synthesis(residual_shifted, lpc)
return wav_shifted.astype(np.int16)
def __test_fftlen(fftlen):
pysptk.c2acr(np.ones(20), 19, fftlen)
def __test_fftlen(fftlen):
pysptk.c2acr(np.ones(20), 19, fftlen)
