def test_gmmmap_swap():
from nnmnkwii.baseline.gmm import MLPG
static_dim = 2
T = 10
windows = _get_windows_set()[-1]
np.random.seed(1234)
src_mc = np.random.rand(T, static_dim * len(windows))
tgt_mc = np.random.rand(T, static_dim * len(windows))
# pseudo parallel data
XY = np.concatenate((src_mc, tgt_mc), axis=-1)
gmm = GaussianMixture(n_components=4)
gmm.fit(XY)
paramgen = MLPG(gmm, windows=windows, swap=False)
swap_paramgen = MLPG(gmm, windows=windows, swap=True)
mc_converted1 = paramgen.transform(src_mc)
mc_converted2 = swap_paramgen.transform(tgt_mc)
src_mc = src_mc[:, :static_dim]
tgt_mc = tgt_mc[:, :static_dim]
assert norm(tgt_mc - mc_converted1) < norm(src_mc - mc_converted1)
assert norm(tgt_mc - mc_converted2) > norm(src_mc - mc_converted2)
def test_diffvc():
from nnmnkwii.baseline.gmm import MLPG
# MLPG is performed dimention by dimention, so static_dim 1 is enough, 2 just for in
# case.
static_dim = 2
T = 10
for windows in _get_windows_set():
np.random.seed(1234)
src_mc = np.random.rand(T, static_dim * len(windows))
tgt_mc = np.random.rand(T, static_dim * len(windows))
# pseudo parallel data
XY = np.concatenate((src_mc, tgt_mc), axis=-1)
gmm = GaussianMixture(n_components=4)
gmm.fit(XY)
paramgen = MLPG(gmm, windows=windows, diff=False)
diff_paramgen = MLPG(gmm, windows=windows, diff=True)
mc_converted1 = paramgen.transform(src_mc)
mc_converted2 = diff_paramgen.transform(src_mc)
assert mc_converted1.shape == (T, static_dim)
assert mc_converted2.shape == (T, static_dim)
src_mc = src_mc[:, :static_dim]
tgt_mc = tgt_mc[:, :static_dim]
assert norm(tgt_mc - mc_converted1) < norm(src_mc - mc_converted1)
def convert(self, feature, mlpg=True, diff=False):
windows = delta.DELTA_WINDOWS
if not mlpg:
windows = windows[0:1]
paramgen = MLPG(self.gmm, windows=windows, diff=diff)
return paramgen.transform(feature)
def transform(self, XY):
X, Y = XY
assert X.ndim == 3 and Y.ndim == 3
longer_features = X if X.shape[1] > Y.shape[1] else Y
Xc = X.copy() # this will be updated iteratively
X_aligned = np.zeros_like(longer_features)
Y_aligned = np.zeros_like(longer_features)
refined_paths = np.empty(len(X), dtype=np.object)
for idx in range(self.n_iter):
for idx, (x, y) in enumerate(zip(Xc, Y)):
x, y = trim_zeros_frames(x), trim_zeros_frames(y)
dist, path = fastdtw(x, y, radius=self.radius, dist=self.dist)
dist /= (len(x) + len(y))
pathx = list(map(lambda l: l[0], path))
pathy = list(map(lambda l: l[1], path))
refined_paths[idx] = pathx
x, y = x[pathx], y[pathy]
max_len = max(len(x), len(y))
if max_len > X_aligned.shape[1] or max_len > Y_aligned.shape[1]:
pad_size = max(max_len - X_aligned.shape[1],
max_len > Y_aligned.shape[1])
X_aligned = np.pad(X_aligned, [(0, 0), (0, pad_size),
(0, 0)],
mode="constant",
constant_values=0)
Y_aligned = np.pad(Y_aligned, [(0, 0), (0, pad_size),
(0, 0)],
mode="constant",
constant_values=0)
X_aligned[idx][:len(x)] = x
Y_aligned[idx][:len(y)] = y
if self.verbose > 0:
print("{}, distance: {}".format(idx, dist))
# Fit
gmm = GaussianMixture(n_components=self.n_components_gmm,
covariance_type="full",
max_iter=self.max_iter_gmm)
XY = np.concatenate((X_aligned, Y_aligned),
axis=-1).reshape(-1, X.shape[-1] * 2)
gmm.fit(XY)
windows = [(0, 0, np.array([1.0]))] # no delta
paramgen = MLPG(gmm, windows=windows)
for idx in range(len(Xc)):
x = trim_zeros_frames(Xc[idx])
Xc[idx][:len(x)] = paramgen.transform(x)
# Finally we can get aligned X
for idx in range(len(X_aligned)):
x = X[idx][refined_paths[idx]]
X_aligned[idx][:len(x)] = x
return X_aligned, Y_aligned
def test_one_utt(src_path, tgt_path, disable_mlpg=False, diffvc=True):
# GMM-based parameter generation is provided by the library in `baseline` module
if disable_mlpg:
# Force disable MLPG
paramgen = MLPG(gmm, windows=[(0, 0, np.array([1.0]))], diff=diffvc)
else:
paramgen = MLPG(gmm, windows=windows, diff=diffvc)
fs, x = wavfile.read(src_path)
x = x.astype(np.float64)
f0, timeaxis = pyworld.dio(x, fs, frame_period=frame_period)
f0 = pyworld.stonemask(x, f0, timeaxis, fs)
spectrogram = pyworld.cheaptrick(x, f0, timeaxis, fs)
aperiodicity = pyworld.d4c(x, f0, timeaxis, fs)
pdb.set_trace()
mc = pysptk.sp2mc(spectrogram, order=order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
if use_delta:
mc = delta_features(mc, windows)
mc = paramgen.transform(mc)
if disable_mlpg and mc.shape[-1] != static_dim:
mc = mc[:, :static_dim]
assert mc.shape[-1] == static_dim
mc = np.hstack((c0[:, None], mc))
if diffvc:
mc[:, 0] = 0 # remove power coefficients
engine = Synthesizer(MLSADF(order=order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
spectrogram = pysptk.mc2sp(mc.astype(np.float64),
alpha=alpha,
fftlen=fftlen)
waveform = pyworld.synthesize(f0, spectrogram, aperiodicity, fs,
frame_period)
return waveform
def test_one_utt(path_src, path_tgt, disable_mlpg=False, diffvc=True):
if disable_mlpg:
paramgen = MLPG(gmm, windows=[(0, 0, np.array([1.0]))], diff=diffvc)
else:
paramgen = MLPG(gmm, windows=windows, diff=diffvc)
x, fs_ = sf.read(path_src)
x = x.astype(np.float64)
f0, time_axis = pyworld.dio(x, fs_, frame_period=frame_period)
f0 = pyworld.stonemask(x, f0, time_axis, fs_)
spectrogram = pyworld.cheaptrick(x, f0, time_axis, fs_)
aperiodicity = pyworld.d4c(x, f0, time_axis, fs_)
mc = pysptk.sp2mc(spectrogram, order=order, alpha=alpha)
c0, mc = mc[:, 0], mc[:, 1:]
if use_delta:
mc = delta_features(mc, windows)
mc = paramgen.transform(mc)
if disable_mlpg and mc.shape[-1] != static_dim:
mc = mc[:, :static_dim]
assert mc.shape[-1] == static_dim
mc = np.hstack((c0[:, None], mc))
if diffvc:
mc[:, 0] = 0
engine = Synthesizer(MLSADF(order=order, alpha=alpha),
hopsize=hop_length)
b = pysptk.mc2b(mc.astype(np.float64), alpha=alpha)
waveform = engine.synthesis(x, b)
else:
spectrogram = pysptk.mc2sp(mc.astype(np.float64),
alpha=alpha,
fftlen=fftlen)
waveform = pyworld.synthesize(f0, spectrogram, aperiodicity, fs_,
frame_period)
return waveform