def test_write_read_int32_vector(tmpdir, endian):
path = tmpdir.mkdir('test')
a = np.random.randint(1, 128, 10, dtype=np.int32)
b = np.random.randint(1, 128, 10, dtype=np.int32)
origin = {u'Ï,é,à': a, u'あいうえお': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('b.scp').strpath,
endian=endian)
d2 = {
k: v
for k, v in kaldiio.load_ark(path.join('a.ark').strpath, endian=endian)
}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join('b.scp').strpath,
endian=endian).items()
}
with io.open(path.join('a.ark').strpath, 'rb') as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd, endian=endian)}
_compare(d2, origin)
_compare(d5, origin)
_compare(d6, origin)
def test_append_mode(tmpdir):
path = tmpdir.mkdir("test")
a = np.random.rand(1000, 120).astype(np.float32)
b = np.random.rand(10, 120).astype(np.float32)
origin = {"Ï,é,à": a, "あいうえお": b}
kaldiio.save_ark(path.join("a.ark").strpath, origin, scp=path.join("b.scp").strpath)
kaldiio.save_ark(
path.join("a2.ark").strpath,
{"Ï,é,à": a},
scp=path.join("b2.scp").strpath,
append=True,
)
kaldiio.save_ark(
path.join("a2.ark").strpath,
{"あいうえお": b},
scp=path.join("b2.scp").strpath,
append=True,
)
d1 = {k: v for k, v in kaldiio.load_ark(path.join("a.ark").strpath)}
d2 = {k: v for k, v in kaldiio.load_scp(path.join("b.scp").strpath).items()}
d3 = {k: v for k, v in kaldiio.load_ark(path.join("a2.ark").strpath)}
d4 = {k: v for k, v in kaldiio.load_scp(path.join("b2.scp").strpath).items()}
_compare(d1, origin)
_compare(d2, origin)
_compare(d3, origin)
_compare(d4, origin)
def test_write_read(tmpdir, shape1, shape2, endian, dtype, max_cache_fd):
path = tmpdir.mkdir("test")
a = np.random.rand(*shape1).astype(dtype)
b = np.random.rand(*shape2).astype(dtype)
origin = {"Ï,é,à": a, "あいうえお": b}
kaldiio.save_ark(
path.join("a.ark").strpath,
origin,
scp=path.join("b.scp").strpath,
endian=endian,
)
d2 = {k: v for k, v in kaldiio.load_ark(path.join("a.ark").strpath, endian=endian)}
d5 = {
k: v
for k, v in kaldiio.load_scp(
path.join("b.scp").strpath, endian=endian, max_cache_fd=max_cache_fd
).items()
}
with io.open(path.join("a.ark").strpath, "rb") as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd, endian=endian)}
_compare(d2, origin)
_compare(d5, origin)
_compare(d6, origin)
def test_write_read_compress(tmpdir, compression_method, endian):
path = tmpdir.mkdir('test')
a = np.random.rand(1000, 120).astype(np.float32)
b = np.random.rand(10, 120).astype(np.float32)
origin = {u'Ï,é,à': a, u'あいうえお': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('b.scp').strpath,
compression_method=compression_method,
endian=endian)
d2 = {
k: v
for k, v in kaldiio.load_ark(path.join('a.ark').strpath, endian=endian)
}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join('b.scp').strpath,
endian=endian).items()
}
with io.open(path.join('a.ark').strpath, 'rb') as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd, endian=endian)}
_compare_allclose(d2, origin, atol=1e-1)
_compare_allclose(d5, origin, atol=1e-1)
_compare_allclose(d6, origin, atol=1e-1)
def test_write_read(tmpdir, shape1, shape2, endian, dtype):
path = tmpdir.mkdir('test')
a = np.random.rand(*shape1).astype(dtype)
b = np.random.rand(*shape2).astype(dtype)
origin = {u'Ï,é,à': a, u'あいうえお': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('b.scp').strpath,
endian=endian)
d2 = {
k: v
for k, v in kaldiio.load_ark(path.join('a.ark').strpath, endian=endian)
}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join('b.scp').strpath,
endian=endian).items()
}
with io.open(path.join('a.ark').strpath, 'rb') as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd, endian=endian)}
_compare(d2, origin)
_compare(d5, origin)
_compare(d6, origin)
def test_append_mode(tmpdir):
path = tmpdir.mkdir('test')
a = np.random.rand(1000, 120).astype(np.float32)
b = np.random.rand(10, 120).astype(np.float32)
origin = {'a': a, 'b': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('b.scp').strpath)
kaldiio.save_ark(path.join('a2.ark').strpath, {'a': a},
scp=path.join('b2.scp').strpath,
append=True)
kaldiio.save_ark(path.join('a2.ark').strpath, {'b': b},
scp=path.join('b2.scp').strpath,
append=True)
d1 = {k: v for k, v in kaldiio.load_ark(path.join('a.ark').strpath)}
d2 = {
k: v
for k, v in kaldiio.load_scp(path.join('b.scp').strpath).items()
}
d3 = {k: v for k, v in kaldiio.load_ark(path.join('a2.ark').strpath)}
d4 = {
k: v
for k, v in kaldiio.load_scp(path.join('b2.scp').strpath).items()
}
_compare(d1, origin)
_compare(d2, origin)
_compare(d3, origin)
_compare(d4, origin)
def test_read_arks(fname):
# Assume arks dir existing at the same directory
ark0 = dict(
kaldiio.load_ark(
os.path.join(os.path.dirname(__file__), 'arks', 'test.ark')))
ark = dict(kaldiio.load_ark(fname))
_compare_allclose(ark, ark0, atol=1e-1)
def test_write_compressed_arks(tmpdir, compression_method):
# Assume arks dir existing at the same directory
ark0 = dict(
kaldiio.load_ark(
os.path.join(os.path.dirname(__file__), 'arks', 'test.ark')))
path = tmpdir.mkdir('test').join('c.ark').strpath
kaldiio.save_ark(path, ark0, compression_method=compression_method)
arkc = dict(kaldiio.load_ark(path))
arkc_valid = dict(
kaldiio.load_ark(
os.path.join(os.path.dirname(__file__), 'arks',
'test.cm{}.ark'.format(compression_method))))
_compare_allclose(arkc, arkc_valid, atol=1e-4)
def _tpl_command(wave: numpy.ndarray,
cmd: Sequence[str],
sample_rate: float,
buffering: int = -1,
verbose: int = -1) -> numpy.ndarray:
if wave.dtype not in (numpy.int8, numpy.int16, numpy.int32):
raise ValueError(
'Can read only PCM data. Input as int8, int16 or int32 type'
'(int24 is not supported by numpy).')
if shutil.which(cmd[0]) is None:
raise RuntimeError(f'Command not found: {cmd[0]}')
with Popen(cmd,
stdin=PIPE,
stdout=PIPE,
stderr=None if verbose > -1 else PIPE,
bufsize=buffering) as p:
with BytesIO() as fin:
scipy.io.wavfile.write(fin, int(sample_rate), wave)
stdin = DUMMY_KEY + fin.getvalue()
stdout, stderr = p.communicate(input=stdin)
if p.returncode != 0:
if stderr is not None:
ms = stderr.decode()
else:
ms = f'Fail: {" ".join(cmd)}'
raise RuntimeError(ms)
fout = BytesIO(stdout)
return next(kaldiio.load_ark(fout))[1]
def KaldiInfer(self, audio):
wav_id = PackZero(self.utt_id, size=6)
self.CreateTestEnv(audio, wav_id)
audio_path = "audio_%s" % PackZero(self.utt_id, size=6)
# pass workspace, infer_set and num of jobs
infer_log = os.popen("%s %s %s 1" % (
os.path.join(self.kaldi_workspace, "extract_post.sh"),
self.kaldi_workspace,
os.path.join(self.kaldi_workspace, "data", audio_path),
))
infer_log = infer_log.readlines()
if "infer success" not in " ".join(infer_log):
print("Error\n%s" % infer_log)
ark_post = os.path.join(self.kaldi_workspace, "data",
audio_path + "_post", "phone_post.1.ark")
post_ark = load_ark(ark_post)
for key, numpy_array in post_ark:
if key == "%s_%s" % (wav_id, wav_id):
post_numpy = numpy_array
break
self.utt_id += 1
# add the column of stress into its non_stress phone
if self.flag_version:
del_list = [
5, 9, 13, 17, 21, 25, 33, 37, 41, 48, 52, 62, 66, 76, 80
]
del_list_all = self.vowels_id
for i in del_list:
for j in [i, i - 1, i - 2]:
post_numpy[:, j - 1] += post_numpy[:, j]
post_numpy = np.delete(post_numpy, del_list_all, axis=1)
return post_numpy
def test_write_read_int32_vector_ascii(tmpdir):
path = tmpdir.mkdir("test")
a = np.random.randint(1, 128, 10, dtype=np.int32)
b = np.random.randint(1, 128, 10, dtype=np.int32)
origin = {"Ï,é,à": a, "あいうえお": b}
kaldiio.save_ark(
path.join("a.ark").strpath, origin, scp=path.join("b.scp").strpath, text=True
)
d2 = {k: v for k, v in kaldiio.load_ark(path.join("a.ark").strpath)}
d5 = {k: v for k, v in kaldiio.load_scp(path.join("b.scp").strpath).items()}
with io.open(path.join("a.ark").strpath, "rb") as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd)}
_compare_allclose(d2, origin)
_compare_allclose(d5, origin)
_compare_allclose(d6, origin)
def test_read_arks():
# Assume arks dir existing at the same directory
arkdir = os.path.join(os.path.dirname(__file__), 'arks')
arks = {
fname: dict(kaldiio.load_ark(fname))
for fname in glob.glob(os.path.join(arkdir, '*.ark'))
}
fnames = list(arks)
for fname in fnames[1:]:
_compare_allclose(arks[fname], arks[fnames[0]], atol=1e-1)
def get_file_dict(self, path: str, uttid_locs: List[Tuple[str, str]] = None) -> \
Dict[str, Dict[str, Any]]:
def mat2dict(mat):
return {"rate": mat[(0, ) * mat.ndim] or None, "x": mat[1:]}
file_dict = OrderedDict(
(uttid, mat2dict(mat)) for uttid, mat in kaldiio.load_ark(path))
if uttid_locs is None:
return file_dict
return OrderedDict((k, file_dict[k]) for k, loc in uttid_locs)
def test_write_load(tmpdir):
path = tmpdir.mkdir('test')
a = np.random.rand(1000, 120).astype(np.float32)
b = np.random.rand(10, 120).astype(np.float32)
origin = {'a': a, 'b': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('b.scp').strpath)
d2 = {k: v for k, v in kaldiio.load_ark(path.join('a.ark').strpath)}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join('b.scp').strpath).items()
}
with open(path.join('a.ark').strpath, 'rb') as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd)}
_compare(d2, origin)
_compare(d5, origin)
_compare(d6, origin)
def test_write_read_ascii(tmpdir):
path = tmpdir.mkdir("test")
a = np.random.rand(10, 10).astype(np.float32)
b = np.random.rand(5, 35).astype(np.float32)
origin = {"Ï,é,à": a, "あいうえお": b}
kaldiio.save_ark(
path.join("a.ark").strpath, origin, scp=path.join("a.scp").strpath, text=True
)
d2 = {k: v for k, v in kaldiio.load_ark(path.join("a.ark").strpath)}
d5 = {k: v for k, v in kaldiio.load_scp(path.join("a.scp").strpath).items()}
_compare_allclose(d2, origin)
_compare_allclose(d5, origin)
def test_write_read_int32_vector_ascii(tmpdir):
path = tmpdir.mkdir('test')
a = np.random.randint(1, 128, 10, dtype=np.int32)
b = np.random.randint(1, 128, 10, dtype=np.int32)
origin = {'a': a, 'b': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('b.scp').strpath,
text=True)
d2 = {k: v for k, v in kaldiio.load_ark(path.join('a.ark').strpath)}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join('b.scp').strpath).items()
}
with open(path.join('a.ark').strpath, 'r') as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd)}
_compare_allclose(d2, origin)
_compare_allclose(d5, origin)
_compare_allclose(d6, origin)
def call(self, stats):
"""
Do CMVN.
:param stats: Statistics of features.
:return: Mean and std of features.
"""
p = self.config
if isinstance(stats, dict):
stats_dict = dict(stats)
else:
if p.filetype == 'mat':
stats_dict = {None: kaldiio.load_mat(stats)}
elif p.filetype == 'ark':
stats_dict = dict(kaldiio.load_ark(stats))
else:
raise ValueError('Not supporting filetype={}'.format(
p.filetype))
if p.utt2spk is not None:
self.utt2spk = {}
with io.open(p.utt2spk, 'r', encoding='utf-8') as f:
for line in f:
utt, spk = line.rstrip().split(None, 1)
self.utt2spk[utt] = spk
elif p.spk2utt is not None:
self.utt2spk = {}
with io.open(p.spk2utt, 'r', encoding='utf-8') as f:
for line in f:
spk, utts = line.rstrip().split(None, 1)
for utt in utts.split():
self.utt2spk[utt] = spk
else:
self.utt2spk = None
self.bias = {}
self.scale = {}
for spk, stats in stats_dict.items():
assert len(stats) == 2, stats.shape
count = stats[0, -1]
if not (np.isscalar(count) or isinstance(count, (int, float))):
count = count.flatten()[0]
mean = stats[0, :-1] / count
var = stats[1, :-1] / count - mean * mean
std = np.maximum(np.sqrt(var), p.std_floor)
self.bias[spk] = -mean
self.scale[spk] = 1 / std
def main():
parser = get_parser()
args = parser.parse_args()
d = kaldiio.load_ark(args.rspecifier)
with file_writer_helper(
args.wspecifier,
filetype='mat',
write_num_frames=args.write_num_frames,
compress=args.compress,
compression_method=args.compression_method) as writer:
for utt, mat in d:
writer[utt] = mat
def test_write_read_compress(tmpdir, compression_method, endian):
path = tmpdir.mkdir("test")
a = np.random.rand(1000, 120).astype(np.float32)
b = np.random.rand(10, 120).astype(np.float32)
origin = {"Ï,é,à": a, "あいうえお": b}
kaldiio.save_ark(
path.join("a.ark").strpath,
origin,
scp=path.join("b.scp").strpath,
compression_method=compression_method,
endian=endian,
)
d2 = {k: v for k, v in kaldiio.load_ark(path.join("a.ark").strpath, endian=endian)}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join("b.scp").strpath, endian=endian).items()
}
with io.open(path.join("a.ark").strpath, "rb") as fd:
d6 = {k: v for k, v in kaldiio.load_ark(fd, endian=endian)}
_compare_allclose(d2, origin, atol=1e-1)
_compare_allclose(d5, origin, atol=1e-1)
_compare_allclose(d6, origin, atol=1e-1)
def test_write_load_ascii(tmpdir):
path = tmpdir.mkdir('test')
a = np.random.rand(10, 10).astype(np.float32)
b = np.random.rand(5, 35).astype(np.float32)
origin = {'a': a, 'b': b}
kaldiio.save_ark(path.join('a.ark').strpath,
origin,
scp=path.join('a.scp').strpath,
text=True)
d2 = {k: v for k, v in kaldiio.load_ark(path.join('a.ark').strpath)}
d5 = {
k: v
for k, v in kaldiio.load_scp(path.join('a.scp').strpath).items()
}
_compare_allclose(d2, origin)
_compare_allclose(d5, origin)
def write_expanded_feature(raw_mfcc_and_pitch_file, output_data_dir):
expanded_features = {}
for utt, features in kaldiio.load_ark(raw_mfcc_and_pitch_file):
num_frames = len(features)
target_emotion_column = np.full((num_frames, 1),
get_target_emotion(utt))
expanded_feature = np.append(features, target_emotion_column, 1)
expanded_features[utt] = expanded_feature
(_, split, _) = raw_mfcc_and_pitch_file.split('.', 2)
kaldiio.save_ark(
os.path.join(output_data_dir,
'mfcc_pitch_and_target_emotion.%s.ark' % (split)),
expanded_features,
scp=os.path.join(output_data_dir,
'mfcc_pitch_and_target_emotion.%s.scp' % (split)))
def KaldiInfer(self, audio):
wav_id = PackZero(self.utt_id, size=6)
self.CreateTestEnv(audio, wav_id)
audio_path = "audio_%s" % PackZero(self.utt_id, size=6)
# pass workspace, infer_set and num of jobs
infer_log = os.popen(
"%s %s %s 1" %
(os.path.join(self.kaldi_workspace,
"extract_post.sh"), self.kaldi_workspace,
os.path.join(self.kaldi_workspace, "data", audio_path)))
infer_log = infer_log.readlines()
if "infer success" not in " ".join(infer_log):
print("Error\n%s" % infer_log)
ark_post = os.path.join(self.kaldi_workspace, "data",
audio_path + "_post", "phone_post.1.ark")
###############
# ark_post = os.path.join('/home/nan/CALL-proto/backend/audio_000151_post_test/phone_post.1.ark')
post_ark = load_ark(ark_post)
for key, numpy_array in post_ark:
if key == "%s_%s" % (wav_id, wav_id):
post_numpy = numpy_array
break
# print(post_numpy.shape)
self.utt_id += 1
### Nan
if self.flag_version:
del_list = [
5, 9, 13, 17, 21, 25, 33, 37, 41, 48, 52, 62, 66, 76, 80
]
del_list_all = [
3, 4, 5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20, 21, 23, 24,
25, 31, 32, 33, 35, 36, 37, 39, 40, 41, 46, 47, 48, 50, 51, 52,
60, 61, 62, 64, 65, 66, 74, 75, 76, 78, 79, 80
]
for i in del_list:
for j in [i, i - 1, i - 2]:
post_numpy[:, j - 1] += post_numpy[:, j]
post_numpy = np.delete(post_numpy, del_list_all, axis=1)
np.savetxt('post_new.csv', post_numpy,
delimiter=',') ### Delete later
###
return post_numpy
def read_cmvn_stats(path, cmvn_type):
if cmvn_type == "global":
d = {None: kaldiio.load_mat(path)}
else:
d = dict(kaldiio.load_ark(path))
return {spk: CMVNStats.from_numpy(stats) for spk, stats in d.items()}
def __init__(self, stats, norm_means=True, norm_vars=False,
filetype='mat', utt2spk=None, spk2utt=None,
reverse=False, std_floor=1.0e-20):
self.stats_file = stats
self.norm_means = norm_means
self.norm_vars = norm_vars
self.reverse = reverse
if isinstance(stats, dict):
stats_dict = dict(stats)
else:
# Use for global CMVN
if filetype == 'mat':
stats_dict = {None: kaldiio.load_mat(stats)}
# Use for global CMVN
elif filetype == 'npy':
stats_dict = {None: np.load(stats)}
# Use for speaker CMVN
elif filetype == 'ark':
self.accept_uttid = True
stats_dict = dict(kaldiio.load_ark(stats))
# Use for speaker CMVN
elif filetype == 'hdf5':
self.accept_uttid = True
stats_dict = h5py.File(stats)
else:
raise ValueError('Not supporting filetype={}'.format(filetype))
if utt2spk is not None:
self.utt2spk = {}
with io.open(utt2spk, 'r', encoding='utf-8') as f:
for line in f:
utt, spk = line.rstrip().split(None, 1)
self.utt2spk[utt] = spk
elif spk2utt is not None:
self.utt2spk = {}
with io.open(spk2utt, 'r', encoding='utf-8') as f:
for line in f:
spk, utts = line.rstrip().split(None, 1)
for utt in utts.split():
self.utt2spk[utt] = spk
else:
self.utt2spk = None
# Kaldi makes a matrix for CMVN which has a shape of (2, feat_dim + 1),
# and the first vector contains the sum of feats and the second is
# the sum of squares. The last value of the first, i.e. stats[0,-1],
# is the number of samples for this statistics.
self.bias = {}
self.scale = {}
for spk, stats in stats_dict.items():
assert len(stats) == 2, stats.shape
count = stats[0, -1]
# If the feature has two or more dimensions
if not (np.isscalar(count) or isinstance(count, (int, float))):
# The first is only used
count = count.flattten()[0]
mean = stats[0, :-1] / count
# V(x) = E(x^2) - (E(x))^2
var = stats[1, :-1] / count - mean * mean
std = np.maximum(np.sqrt(var), std_floor)
self.bias[spk] = -mean
self.scale[spk] = 1 / std
def __init__(self, norm_means=True, norm_vars=False, std_floor=1.0e-20):
self.norm_means = norm_means
self.norm_vars = norm_vars
self.std_floor = std_floor
def __repr__(self):
return ('{name}(norm_means={norm_means}, norm_vars={norm_vars})'
.format(name=self.__class__.__name__,
norm_means=self.norm_means, norm_vars=self.norm_vars))
def __call__(self, x, uttid=None):
# x: [Time, Dim]
square_sums = (x ** 2).sum(axis=0)
mean = x.mean(axis=0)
if self.norm_means:
x = np.subtract(x, mean)
if self.norm_vars:
var = square_sums / x.shape[0] - mean ** 2
std = np.maximum(np.sqrt(var), self.std_floor)
x = np.divide(x, std)
return x
if __name__ == '__main__':
a = kaldiio.load_ark("/teamscratch/tts_intern_experiment/yuwu1/ASR/script/cmvn.ark")
#for key, numpy_array in a:
# print(key,numpy_array)
#stats_dict = dict()
print(dict(a))
def main(argv):
""" Main script
:param argv: train or test path
:returns: None
"""
feature_name = 'energy'
escale_16_PCM = 2**15
#import soundfile as sf
#ob = sf.SoundFile(os.path.join(audio_path,kal[:-3]+'wav'))
#print('Sample rate: {}'.format(ob.samplerate))
#print('Channels: {}'.format(ob.channels))
#print('Subtype: {}'.format(ob.subtype))
#16-bit PCM -32768 +32767 int16
# Se define el path de test o train
path = sys.argv[1]
folder = path.split("/")[1]
# folder = 'train'
kal_list = os.listdir(os.path.join('info_user', folder))
# 1.- Cargar el archivo de configuración
frame_shift, frame_length = load_config()
# 2.- Cargar los audios y dividirlos en ventanas para calcular la energía
audio_path = 'audio/experiment_lm'
energy_dict = {}
# t = time.time()
for kal in kal_list:
if kal.endswith('.kal'):
# 3.- Calcula la energía y las derivadas
fs, audio = wavfile.read(os.path.join(audio_path,
kal[:-3] + 'wav'))
audio = audio / escale_16_PCM
kal_file = open(os.path.join('info_user', folder, kal), 'r')
kal_file = kal_file.readlines()
energy_vector, deltae_vector, deltade_vector = calculate_energy(
audio, fs, frame_shift, frame_length, kal_file)
# Este ajuste se hace porque al generar el feature el orden es 1,2,3,... pero al guardarlos tenemos que tener 1,10,11,12,... porque es el orden que utiliza kaldi y los diccionarios.
key_list = list()
for i in range(0, len(energy_vector)):
key_list.append(kal[:-4] + '-utt' + str(i))
# end for
key_list.sort()
for i in range(0, len(energy_vector)):
index = int(key_list[i][16:])
#print(key_list[i])
#print(index)
energy_dict[key_list[i]] = np.vstack(
(np.asarray(energy_vector[index]),
np.asarray(deltae_vector[index]),
np.asarray(deltade_vector[index]))).T
# end for
# end if
# end for
# elapsed = time.time() - t
# 4.- Lo guarda como un .ark y .scp
if os.path.exists('mfcc') and len(os.listdir('mfcc')) > 0:
features_listdir = os.listdir('mfcc')
folder_features = 'mfcc'
elif os.path.exists('plp') and len(os.listdir('plp')) > 0:
features_listdir = os.listdir('plp')
folder_features = 'plp'
else:
# No existen carpeta sgeneradas por kaldi para los features
features_listdir = list()
folder_features = 'none'
# end if
if folder_features == 'none':
# Se cogen los features calculados y se dividen en 4. Creamos nuestro propio listdir
d = split_dict(energy_dict, 4)
index = 1
for dic in d:
destark_filename = 'raw_' + feature_name + '_' + folder + '.' + str(
index) + '.ark'
destark_filename = os.path.join(os.getcwd(), feature_name,
destark_filename)
srcscp_filename = destark_filename.replace('ark', 'scp')
#print ("Writing to " + destark_filename)
kaldiio.save_ark(destark_filename, dic, scp=srcscp_filename)
index = index + 1
# end for
else:
for file in features_listdir:
write_dict = {} # kaldiio uses features in the form of a dict
if file.endswith('.ark') and folder in file:
d = kaldiio.load_ark(os.path.join(folder_features, file))
for key_kaldi, array_kaldi in d:
write_dict[key_kaldi] = energy_dict[key_kaldi]
# end for
destark_filename = file.replace(folder_features, feature_name)
destark_filename = os.path.join(os.getcwd(), feature_name,
destark_filename)
srcscp_filename = destark_filename.replace('ark', 'scp')
#print ("Writing to " + destark_filename)
kaldiio.save_ark(destark_filename,
write_dict,
scp=srcscp_filename)
def main(folder):
print("Relocating features...")
# folder = 'train'
if folder == 'train':
featuresCombined_path = 'combined_features'
infouser_path = 'info_user'
ark_list = os.listdir(featuresCombined_path)
# 1.- Load .ark data and save in global dict
global_dict = {}
for ark in ark_list:
if 'raw_data' in ark and ark.endswith('.ark'):
d = kaldiio.load_ark(os.path.join(featuresCombined_path, ark))
for key_kaldi, array_kaldi in d:
global_dict[key_kaldi] = array_kaldi
# end if
# end for
# 2.- Generate train and test dict
kal_train = os.listdir(os.path.join(infouser_path, 'train'))
kal_train = [k[:-4] for k in kal_train] # -4 delete .kal string
kal_train.sort()
kal_test = os.listdir(os.path.join(infouser_path, 'test'))
kal_test = [k[:-4] for k in kal_test]
kal_test.sort()
train_dict = {}
test_dict = {}
global_list = list(global_dict.keys())
global_list.sort()
for key_kaldi in global_list:
if key_kaldi[:-5] in kal_train:
train_dict[key_kaldi] = global_dict[key_kaldi]
elif key_kaldi[:-5] in kal_test:
test_dict[key_kaldi] = global_dict[key_kaldi]
# end if
# end for
# 3.- Save both dictionary as .ark files
for f in ark_list:
os.remove(os.path.join(featuresCombined_path, f))
num_split = int(
len([s for s in ark_list if ('data_' + folder) in str(s)]) / 2)
split_train = split_dict(train_dict, num_split)
split_test = split_dict(test_dict, num_split)
index = 1
for dic in split_train:
destark_filename = 'raw_data_train.' + str(index) + '.ark'
destark_filename = os.path.join(os.getcwd(), featuresCombined_path,
destark_filename)
srcscp_filename = destark_filename.replace('ark', 'scp')
kaldiio.save_ark(destark_filename, dic, scp=srcscp_filename)
index = index + 1
# end for
index = 1
for dic in split_test:
destark_filename = 'raw_data_test.' + str(index) + '.ark'
destark_filename = os.path.join(os.getcwd(), featuresCombined_path,
destark_filename)
srcscp_filename = destark_filename.replace('ark', 'scp')
kaldiio.save_ark(destark_filename, dic, scp=srcscp_filename)
index = index + 1
# end for
# end if
# 4.- Calculate rest of makeFeats.sh
bashCommand = "bash local/copy_scp_all.sh " + folder
process = subprocess.Popen(bashCommand, shell=True)
output, error = process.communicate()
bashCommand = "bash utils/data/fix_data_dir.sh data/" + folder
process = subprocess.Popen(bashCommand, shell=True)
output, error = process.communicate()
bashCommand = "steps/compute_cmvn_stats.sh data/" + folder + " exp/make_combined_features/" + folder + " combined_features || exit 1;"
process = subprocess.Popen(bashCommand, shell=True)
output, error = process.communicate()
bashCommand = "utils/validate_data_dir.sh data/" + folder
process = subprocess.Popen(bashCommand, shell=True)
output, error = process.communicate()
print("Features relocated.")
def align(
array: numpy.ndarray,
transcription: Union[Sequence[int], numpy.ndarray],
tree: str,
transition_model: str,
fst: str,
batch_size: int=250,
read_disambig_syms: str=None,
reorder: bool=True,
rm_eps: bool=False,
transition_scale=1.0,
acoustic_scale: float=1,
self_loop_scale=0.1,
beam: float=200,
binary: bool=True,
careful: bool=False,
retry_beam: float=0,
verbose: int=-1) -> numpy.ndarray:
"""
compile-train-graph + align-compiled-mapped
Args:
array (numpy.ndarray):
transcription (Sequence[int]): A sequnce of integer representing
the transcription of a utterance.
It can be get by this command,
% utils/sym2int.pl --map-oov $(cat $lang/oov.int)
-f 2- $lang/words.txt $text |"
tree (str): tree-in
transition_model (str): transition-model-in("final.mdl")
fst (str): lexicon-fst-in("$lang/L.fst")
batch_size (int): Number of FSTs to compile at a time
(more -> faster but uses more memory. E.g. 500
read_disambig_syms (str): File containing list of disambiguation
symbols in phone symbol table, $lang/phones/disambig.int
reorder (bool): Reorder transition ids for greater decoding
efficiency.
rm_eps (bool) : Remove [most] epsilons before minimization
(only applicable if disambig symbols present)
transition_scale (float):
acoustic_scale (float): Scaling factor for acoustic likelihoods
self_loop_scale (float):
beam (float): Decoding beam used in alignment
binary (bool): Write output in binary mode
careful (bool): If true, do 'careful' alignment,
which is better at detecting alignment failure
(involves loop to start of decoding graph).
retry_beam (float): Decoding beam for second try at alignment
verbose (int): Verbose level (higher->more logging)
"""
for cmd in ['compile-train-graphs', 'align-compiled-mapped']:
if shutil.which(cmd) is None:
raise RuntimeError(
f'Command not found: {cmd}')
if isinstance(transcription, numpy.ndarray) and \
len(transcription.shape) == 1 and transcription.dtype.kind == 'i':
transcription = [int(v) for v in transcription]
if not isinstance(transcription, collections.abc.Sequence) or\
any(not isinstance(v, int) for v in transcription):
raise TypeError(f'{type(transcription)} is not a sequence of integer')
if read_disambig_syms is None:
read_disambig_syms = ''
for f in [tree, transition_model, fst] +\
[read_disambig_syms] if read_disambig_syms != '' else []:
if not Path(f).exists():
raise FileNotFoundError(f'{f}: No such file or directory')
if array.ndim != 2:
raise ValueError(
'The input posterior must be 2-dimension array: '
f'{array.ndim} != 2')
if len(array) < len(transcription):
raise ValueError(
f'The frame length of the input likehood array '
f'is shorter than the length of the transcription. '
f'{len(array)} < {len(transcription)}')
# Don't support self-loop-scale and transition-scale
# because I don't understand how to input...
compile_train_graph = (
f'echo {DUMMY_KEY} {" ".join(map(str, transcription))} | '
f'compile-train-graphs '
f'--batch-size={batch_size} '
f'--read-disambig-syms={read_disambig_syms} '
f'--reorder={str(reorder).lower()} '
f'--rm-eps={str(rm_eps).lower()} '
f'--verbose={max(verbose, 0)} '
f'{tree} {transition_model} {fst} ark:- ark:- |')
cmds = ['align-compiled-mapped',
f'--transition-scale={transition_scale}',
f'--self-loop-scale={self_loop_scale}',
f'--acoustic-scale={acoustic_scale}',
f'--beam={beam}',
f'--binary={str(binary).lower()}',
f'--careful={str(careful).lower()}',
f'--retry-beam={retry_beam}',
f'--verbose={max(verbose, 0)}',
f'{transition_model}',
f'ark:{compile_train_graph}', 'ark:-', 'ark:-']
with Popen(cmds, stdin=PIPE, stdout=PIPE,
stderr=None if verbose > -1 else PIPE,
bufsize=-1) as p:
try:
kaldiio.save_ark(p.stdin, {DUMMY_KEY: array})
except BrokenPipeError:
pass # communicate() must ignore broken pipe errors.
except OSError as e:
if e.errno == errno.EINVAL and p.poll() is not None:
# Issue #19612: On Windows, stdin.write() fails with EINVAL
# if the process already exited before the write
pass
else:
raise
stdout, stderr = p.communicate()
if p.returncode != 0:
if stderr is not None:
ms = stderr.decode()
else:
ms = f'Fail: {" ".join(cmds)}'
raise RuntimeError(ms)
fout = BytesIO(stdout)
return next(kaldiio.load_ark(fout))[1]
ark_dir = ['trainNet_sp/deltafalse', 'valid/deltafalse', 'test/deltafalse']
ark_dir = [data_dir+x for x in ark_dir]
save_dir = 'data/timit_dataset/'
save_sub_dirs = ['train/', 'valid/', 'test/']
save_sub_dirs = [save_dir+x for x in save_sub_dirs]
if not os.path.exists(save_dir):
os.makedirs(save_dir)
for i in save_sub_dirs:
if not os.path.exists(i):
os.makedirs(i)
for i, data_type in enumerate(ark_dir):
dir = data_type
save_path = save_sub_dirs[i]
label_file = label_files[i]
df = pd.read_csv(label_file, sep='\t', header=None)
df.set_index(0, inplace=True)
for file in glob(f'{dir}/*.ark'):
d = kaldiio.load_ark(file)
for key, numpy_array in d:
label = df[1][key]
print(f'{save_path}{key}_{label}.npy')
np.save(f'{save_path}{key}_{label}.npy', numpy_array.T)
for i in save_sub_dirs:
datatype = i.split('/')[1]
num = len(os.listdir(i))
print(f'{datatype} => {num}')
#x-vectors from the .ark files for single speakers
import numpy as np
import kaldiio
import os
for filename in os.listdir('espnet/egs/libritts/tts1/exp/xvector_nnet_1a/xvectors_train_clean_100'):
if filename.endswith('.ark'):
f = kaldiio.load_ark(os.path.join('espnet/egs/libritts/tts1/exp/xvector_nnet_1a/xvectors_train_clean_100', filename)) # open in readonly mode
for key, numpy_array in f:
print(filename,key)
path = 'xvector/'+key+'.txt'
print(path)
np.savetxt(path, numpy_array)
