def test_dtw_aligner():
from nnmnkwii.preprocessing.alignment import DTWAligner, IterativeDTWAligner
x, fs = librosa.load(example_audio_file(), sr=None)
assert fs == 16000
x_fast = librosa.effects.time_stretch(x, 2.0)
X = _get_mcep(x, fs)
Y = _get_mcep(x_fast, fs)
D = X.shape[-1]
# Create padded pair
X, Y = adjust_frame_lengths(X, Y, divisible_by=2)
# Add utterance axis
X = X.reshape(1, -1, D)
Y = Y.reshape(1, -1, D)
X_aligned, Y_aligned = DTWAligner().transform((X, Y))
assert X_aligned.shape == Y_aligned.shape
assert np.linalg.norm(X_aligned - Y_aligned) < np.linalg.norm(X - Y)
X_aligned, Y_aligned = IterativeDTWAligner(n_iter=2,
max_iter_gmm=10,
n_components_gmm=2).transform(
(X, Y))
assert X_aligned.shape == Y_aligned.shape
assert np.linalg.norm(X_aligned - Y_aligned) < np.linalg.norm(X - Y)
# Custom dist function
from nnmnkwii.metrics import melcd
X_aligned, Y_aligned = DTWAligner(dist=melcd).transform((X, Y))
assert np.linalg.norm(X_aligned - Y_aligned) < np.linalg.norm(X - Y)
def test_adjust_frame_lengths():
T1 = 10
T2 = 11
D = 5
# 1d and 2d padding
for (x, y) in [(np.random.rand(T1), np.random.rand(T2)),
(np.random.rand(T1, D), np.random.rand(T2, D))]:
x_hat, y_hat = adjust_frame_lengths(x, y, pad=True)
assert x_hat.shape == y_hat.shape
assert x_hat.shape[0] == 11
x_hat, y_hat = adjust_frame_lengths(x, y, pad=False)
assert x_hat.shape == y_hat.shape
assert x_hat.shape[0] == 10
x_hat, y_hat = adjust_frame_lengths(x, y, pad=True, divisible_by=2)
assert x_hat.shape == y_hat.shape
assert x_hat.shape[0] == 12
x_hat, y_hat = adjust_frame_lengths(x, y, pad=False, divisible_by=2)
assert x_hat.shape == y_hat.shape
assert x_hat.shape[0] == 10
# Divisible
x_hat, y_hat = adjust_frame_lengths(x, y, pad=False, divisible_by=3)
assert x_hat.shape == y_hat.shape
assert x_hat.shape[0] == 9
x_hat, y_hat = adjust_frame_lengths(x, y, pad=True, divisible_by=3)
assert x_hat.shape == y_hat.shape
assert x_hat.shape[0] == 12
# make sure we do zero padding
x = np.random.rand(T1)
y = np.random.rand(T2)
x_hat, y_hat = adjust_frame_lengths(x, y, pad=True, divisible_by=3)
assert x_hat[-1] == 0 and y_hat[-1] == 0
# make sure we passes extra kwargs to np.pad
x_hat, y_hat = adjust_frame_lengths(x,
y,
pad=True,
divisible_by=3,
mode="constant",
constant_values=1)
assert x_hat[-1] == 1 and y_hat[-1] == 1
print("Destination dir for {}: {}".format(speaker, d))
if not exists(d):
os.makedirs(d)
# Convert to arrays
print("Convert datasets to arrays")
X, Y = X_dataset.asarray(verbose=1), Y_dataset.asarray(verbose=1)
# Alignment
print("Perform alignment")
X, Y = DTWAligner().transform((X, Y))
print("Save features to disk")
for idx, (x, y) in tqdm(enumerate(zip(X, Y))):
# paths
src_name = splitext(basename(X_dataset.collected_files[idx][0]))[0]
tgt_name = splitext(basename(Y_dataset.collected_files[idx][0]))[0]
src_path = join(dst_dir, "X", src_name)
tgt_path = join(dst_dir, "Y", tgt_name)
# Trim and ajast frames
x = P.trim_zeros_frames(x)
y = P.trim_zeros_frames(y)
x, y = P.adjust_frame_lengths(x, y, pad=True, divisible_by=2)
# Save
np.save(src_path, x)
np.save(tgt_path, y)
sys.exit(0)
print("Destination dir for {}: {}".format(speaker, d))
if not exists(d):
os.makedirs(d)
# Convert to arrays
print("Convert datasets to arrays")
X, Y = X_dataset.asarray(verbose=1), Y_dataset.asarray(verbose=1)
# Alignment
print("Perform alignment")
X, Y = DTWAligner().transform((X, Y))
print("Save features to disk")
for idx, (x, y) in tqdm(enumerate(zip(X, Y))):
# paths
src_name = splitext(basename(X_dataset.collected_files[idx][0]))[0]
tgt_name = splitext(basename(Y_dataset.collected_files[idx][0]))[0]
src_path = join(dst_dir, "X", src_name)
tgt_path = join(dst_dir, "Y", tgt_name)
# Trim and ajast frames
x = P.trim_zeros_frames(x)
y = P.trim_zeros_frames(y)
x, y = P.adjust_frame_lengths(x, y, pad=True, divisible_by=2)
# Save
np.save(src_path, x)
np.save(tgt_path, y)
sys.exit(0)
