def test_energy(audio, mfcc):
# VAD from mfcc or energy is the same excepted properties
vad1 = VadPostProcessor().process(EnergyProcessor().process(audio))
vad2 = VadPostProcessor().process(mfcc)
vad1 = Features(vad1.data, vad1.times)
vad2 = Features(vad2.data, vad2.times)
assert vad1 == vad2
def test_tofrom_dict(mfcc):
a = mfcc._to_dict()
b = Features._from_dict(a)
assert b == mfcc
with pytest.raises(ValueError) as err:
Features._from_dict({'data': a['data'], 'properties': a['properties']})
assert 'missing keys: times' in str(err)
def test_1d_times_sorted():
# 10 frames, 5 dims
data = np.random.random((10, 5))
p = MfccProcessor()
times = p.times(10)
assert times.shape == (10, 2)
feats = Features(data, times[:, 1], validate=False)
assert feats.is_valid()
def test_collection_isclose():
f1 = Features(np.random.random((10, 2)), np.ones((10, )))
f2 = Features(np.random.random((10, 2)), np.ones((10, )))
fc1 = FeaturesCollection(f1=f1, f2=f2)
fc2 = FeaturesCollection(f1=f1, f2=Features(f2.data + 1, f2.times))
fc3 = FeaturesCollection(f1=f1, f3=f2)
assert fc1.is_close(fc1)
assert not fc1.is_close(fc2)
assert fc1.is_close(fc2, atol=1)
assert not fc1.is_close(fc3)
def mfcc_utf8(mfcc):
props = mfcc.properties
props['comments'] = '使用人口について正確な統計はないが、日本国'
feats = FeaturesCollection()
feats['æðÐ'] = Features(mfcc.data, mfcc.times, props)
return feats
def process(self, features):
"""Computes voice activity detection (VAD) on the input `features`
Parameters
----------
features : :class:`~shennong.features.features.Features`, shape = [n,m]
The speech features on which to look for voiced
frames. The first coefficient must be a log-energy (or
equivalent). Works well with
:class:`~shennong.features.processor.mfcc.MfccProcessor` and
:class:`~shennong.features.processor.plp.PlpProcessor`.
Returns
-------
vad : :class:`~shennong.features.features.Features`, shape = [n,1]
The output vad features are of dtype uint8 and contain 1
for voiced frames or 0 for unvoiced frames.
"""
data = kaldi.matrix.SubVector(
kaldi.ivector.compute_vad_energy(
self._options, kaldi.matrix.SubMatrix(features.data))).numpy()
return Features(np.atleast_2d(data.astype(np.uint8)).T,
features.times,
properties=self.get_properties(features))
def _process(self, cls, signal, vtln_warp):
"""Inner process method common to all Kaldi Mel processors"""
# ensure the signal is correct
if signal.nchannels != 1:
raise ValueError(
'signal must have one dimension, but it has {}'.format(
signal.nchannels))
if self.sample_rate != signal.sample_rate:
raise ValueError('processor and signal mismatch in sample rates: '
'{} != {}'.format(self.sample_rate,
signal.sample_rate))
# we need to forward options (because the assignation here is
# done by copy, not by reference. If the user do 'p =
# Processor(); p.dither = 0', this is forwarded to Kaldi here)
self._options.frame_opts = self._frame_options
self._options.mel_opts = self._mel_options
# force 16 bits integers
signal = signal.astype(np.int16).data
data = kaldi.matrix.SubMatrix(
cls(self._options).compute(kaldi.matrix.SubVector(signal),
vtln_warp)).numpy()
return Features(data,
self.times(data.shape[0]),
properties=self.get_properties())
def test_save_invalid(tmpdir, mfcc):
f = str(tmpdir.join('foo.json'))
h = serializers.get_serializer(FeaturesCollection, f, None)
feats = FeaturesCollection(
mfcc=Features(data=mfcc.data, times=0, validate=False))
with pytest.raises(ValueError) as err:
h.save(feats)
assert 'features are not valid' in str(err.value)
def test_apply_baddim(features_collection):
feats = FeaturesCollection(
{k: v.copy() for k, v in features_collection.items()})
feats['new'] = Features(
np.random.random((2, 1)), np.asarray([0, 1]))
with pytest.raises(ValueError) as err:
apply_cmvn(feats)
assert 'must have consistent dimensions' in str(err.value)
def features_collection():
# build a collection of 3 random features of same ndims, various
# nframes
dim = 10
feats = FeaturesCollection()
for n in range(3):
nframes = np.random.randint(5, 15)
feats[str(n)] = Features(np.random.random((nframes, dim)),
np.arange(0, nframes))
return feats
def test_partition():
f1 = Features(np.random.random((10, 2)), np.ones((10, )))
f2 = Features(np.random.random((5, 2)), np.ones((5, )))
f3 = Features(np.random.random((5, 2)), np.ones((5, )))
fc = FeaturesCollection(f1=f1, f2=f2, f3=f3)
with pytest.raises(ValueError) as err:
fp = fc.partition({'f1': 'p1', 'f2': 'p1'})
assert ('following items are not defined in the partition index: f3'
in str(err))
fp = fc.partition({'f1': 'p1', 'f2': 'p1', 'f3': 'p2'})
assert sorted(fp.keys()) == ['p1', 'p2']
assert sorted(fp['p1'].keys()) == ['f1', 'f2']
assert sorted(fp['p2'].keys()) == ['f3']
assert fc.is_valid()
for fc in fp.values():
assert fc.is_valid()
def process(self, signal, vtln_warp=1.0):
"""Compute spectrogram with the specified options
Do an optional feature-level vocal tract length normalization
(VTLN) when `vtln_warp` != 1.0.
Parameters
----------
signal : Audio, shape = [nsamples, 1]
The input audio signal to compute the features on, must be
mono
vtln_warp : float, optional
The VTLN warping factor to be applied when computing
features. Be 1.0 by default, meaning no warping is to be
done.
Returns
-------
features : `Features`, shape = [nframes, `ndims`]
The computed features, output will have as many rows as there
are frames (depends on the specified options `frame_shift`
and `frame_length`).
Raises
------
ValueError
If the input `signal` has more than one channel (i.e. is
not mono). If `sample_rate` != `signal.sample_rate`.
"""
# ensure the signal is correct
if signal.nchannels != 1:
raise ValueError(
'signal must have one dimension, but it has {}'.format(
signal.nchannels))
if self.sample_rate != signal.sample_rate:
raise ValueError('processor and signal mismatch in sample rates: '
'{} != {}'.format(self.sample_rate,
signal.sample_rate))
# we need to forward options (because the assignation here is
# done by copy, not by reference. If the user do 'p =
# Processor(); p.dither = 0', this is forwarded to Kaldi here)
self._options.frame_opts = self._frame_options
# force 16 bits integers
signal = signal.astype(np.int16).data
data = kaldi.matrix.SubMatrix(
kaldi.feat.spectrogram.Spectrogram(self._options).compute(
kaldi.matrix.SubVector(signal), vtln_warp)).numpy()
return Features(data,
self.times(data.shape[0]),
properties=self.get_properties())
def test_post_pitch(raw_pitch):
post_processor = PitchPostProcessor()
params = post_processor.get_params()
data = post_processor.process(raw_pitch)
assert data.shape[1] == 3
assert raw_pitch.shape[0] == data.shape[0]
assert np.array_equal(raw_pitch.times, data.times)
assert params == post_processor.get_params()
bad_pitch = Features(np.random.random((raw_pitch.nframes, 1)),
raw_pitch.times)
with pytest.raises(ValueError) as err:
post_processor.process(bad_pitch)
assert 'data shape must be (_, 2), but it is (_, 1)' in str(err.value)
bad_pitch = Features(np.random.random((raw_pitch.nframes, 3)),
raw_pitch.times)
with pytest.raises(ValueError) as err:
post_processor.process(bad_pitch)
assert 'data shape must be (_, 2), but it is (_, 3)' in str(err.value)
def test_concatenate_tolerance(capsys):
get_logger(level='info')
f1 = Features(np.random.random((12, 2)), np.ones((12, )))
f2 = Features(np.random.random((10, 2)), np.ones((10, )))
with pytest.raises(ValueError) as err:
f1.concatenate(f2, tolerance=0)
assert 'features have a different number of frames' in str(err)
with pytest.raises(ValueError) as err:
f1.concatenate(f2, tolerance=1)
assert 'features differs number of frames, and greater than ' in str(err)
f3 = f1.concatenate(f2, tolerance=2)
assert f3.shape == (10, 4)
assert 'WARNING' in capsys.readouterr().err
f3 = f2.concatenate(f1, tolerance=2)
assert f3.shape == (10, 4)
assert 'WARNING' in capsys.readouterr().err
def test_concatenate(mfcc):
mfcc2 = mfcc.concatenate(mfcc)
assert mfcc2.nframes == mfcc.nframes
assert mfcc2.ndims == mfcc.ndims * 2
assert mfcc2.properties != mfcc.properties
assert mfcc2.properties['mfcc'] == mfcc.properties['mfcc']
mfcc2 = Features(mfcc.data, mfcc.times + 1)
with pytest.raises(ValueError) as err:
mfcc.concatenate(mfcc2)
assert 'times are not equal' in str(err)
def test_init_bad():
with pytest.raises(ValueError) as err:
Features(0, 0, properties=0)
assert 'data must be a numpy array' in str(err)
with pytest.raises(ValueError) as err:
Features(np.asarray([0]), 0, properties=0)
assert 'times must be a numpy array' in str(err)
with pytest.raises(ValueError) as err:
Features(np.asarray([0]), np.asarray([0]), properties=0)
assert 'properties must be a dictionnary' in str(err)
with pytest.raises(ValueError) as err:
Features(np.asarray([0]), np.asarray([0]), properties={0: 0})
assert 'data dimension must be 2' in str(err)
with pytest.raises(ValueError) as err:
Features(np.asarray([[0], [0]]), np.random.random((2, 2, 2)))
assert 'times dimension must be 1 or 2' in str(err)
with pytest.raises(ValueError) as err:
data = np.random.random((12, 2))
data[2, 1] = np.nan
Features(data, np.ones((12, )))
assert 'data contains non-finite numbers' in str(err)
def process(self, alignment):
# build a bijection token <-> onehot index
token2index = self._token2index(alignment)
# sample the alignment at the requested sample rate
sampled = alignment.at_sample_rate(self.frame.sample_rate)
# get the frames as pairs (istart:istop)
nframes = self.frame.nframes(sampled.shape[0])
frame_boundaries = self.frame.boundaries(nframes)
# allocate the features data
data = np.zeros((frame_boundaries.shape[0], len(token2index)),
dtype=np.bool)
# allocate the window function
window = shennong.features.window.window(
self.frame.samples_per_frame,
type=self.window_type,
blackman_coeff=self.blackman_coeff)
for i, (onset, offset) in enumerate(frame_boundaries):
framed = sampled[onset:offset]
# the frame is made of a single token, no needs to compute
# a window function
if np.all(framed[0] == framed[1:]):
winner = framed[0]
else:
# several tokens in the frame, compute the weights
weights = collections.defaultdict(int)
for j, w in enumerate(window):
weights[framed[j]] += w
# the winner token has the biggest weight
winner = sorted(weights.items(),
key=operator.itemgetter(1),
reverse=True)[0][0]
data[i, token2index[winner]] = 1
try:
properties = self.get_properties()
except ValueError: # tokens not defined
self.tokens = token2index.keys()
properties = self.get_properties()
self.tokens = None
properties[self.name].update({'token2index': token2index})
return Features(data,
frame_boundaries / self.frame.sample_rate,
properties=properties)
def test_times_1d(serializer, tmpdir):
filename = ('feats.ark'
if serializer is serializers.KaldiSerializer else 'feats')
tmpfile = str(tmpdir.join(filename))
p = MfccProcessor()
times = p.times(10)[:, 1]
assert times.shape == (10, )
col = FeaturesCollection(mfcc=Features(np.random.random((10, 5)), times))
serializer(col.__class__, tmpfile).save(col)
col2 = serializer(col.__class__, tmpfile).load()
assert col == col2
def test_equal(mfcc):
# same object
assert mfcc == mfcc
assert mfcc.is_close(mfcc)
# same data
mfcc2 = mfcc.copy()
assert mfcc == mfcc2
assert mfcc.is_close(mfcc2)
# not same shape
mfcc2 = mfcc.concatenate(mfcc)
assert not mfcc == mfcc2
assert not mfcc.is_close(mfcc2)
# not same dtype
mfcc64 = mfcc.copy(dtype=np.float64)
assert not mfcc == mfcc64
assert mfcc.is_close(mfcc64)
# not same properties
mfcc2 = Features(mfcc.data, mfcc.times, properties={'foo': 0})
assert not mfcc == mfcc2
assert not mfcc.is_close(mfcc2)
# not same times
mfcc2 = Features(mfcc.data, mfcc.times + 1, properties=mfcc.properties)
assert not mfcc == mfcc2
assert not mfcc.is_close(mfcc2)
# not same data
mfcc2 = Features(mfcc.data + 1, mfcc.times, properties=mfcc.properties)
assert not mfcc == mfcc2
assert not mfcc.is_close(mfcc2)
# not same data but close
mfcc2 = Features(mfcc.data + 1, mfcc.times, properties=mfcc.properties)
assert not mfcc == mfcc2
assert mfcc.is_close(mfcc2, atol=1)
# not same times but close
mfcc2 = Features(mfcc.data, mfcc.times + 1, properties=mfcc.properties)
assert not mfcc == mfcc2
assert not mfcc.is_close(mfcc2, atol=1)
def process(self, raw_pitch):
"""Post process a raw pitch data as specified by the options
Parameters
----------
raw_pitch : Features, shape = [n, 2]
The pitch as extracted by the `PitchProcessor.process`
method
Returns
-------
pitch : Features, shape = [n, 1 2 3 or 4]
The post-processed pitch usable as speech features. The
output columns are 'pov_feature', 'normalized_log_pitch',
delta_pitch' and 'raw_log_pitch', in that order,if their
respective options are set to True.
Raises
------
ValueError
If `raw_pitch` has not exactly two columns. If all the
following options are False: 'add_pov_feature',
'add_normalized_log_pitch', 'add_delta_pitch' and
'add_raw_log_pitch' (at least one of them must be True).
"""
# check at least one required option is True
if not (self.add_pov_feature or self.add_normalized_log_pitch
or self.add_delta_pitch or self.add_raw_log_pitch):
raise ValueError(
'at least one of the following options must be True: '
'add_pov_feature, add_normalized_log_pitch, '
'add_delta_pitch, add_raw_log_pitch')
if raw_pitch.shape[1] != 2:
raise ValueError(
'data shape must be (_, 2), but it is (_, {})'
.format(raw_pitch.shape[1]))
data = kaldi.matrix.SubMatrix(
kaldi.feat.pitch.process_pitch(
self._options, kaldi.matrix.SubMatrix(raw_pitch.data))).numpy()
return Features(
data, raw_pitch.times, properties=self.get_properties(raw_pitch))
def process(self, signal):
"""Extracts the (NCCF, pitch) from a given speech `signal`
Parameters
----------
signal : Audio
The speech signal on which to estimate the pitch. The
signal's sample rate must match the sample rate specified
in the `PitchProcessor` options.
Returns
-------
raw_pitch_features : Features, shape = [nframes, 2]
The output array has as many rows as there are frames
(depends on the specified options `frame_shift` and
`frame_length`), and two columns corresponding to (NCCF,
pitch).
Raises
------
ValueError
If the input `signal` has more than one channel (i.e. is
not mono). If `sample_rate` != `signal.sample_rate`.
"""
if signal.nchannels != 1:
raise ValueError(
'audio signal must have one channel, but it has {}'
.format(signal.nchannels))
if self.sample_rate != signal.sample_rate:
raise ValueError(
'processor and signal mismatch in sample rates: '
'{} != {}'.format(self.sample_rate, signal.sample_rate))
# force 16 bits integers
signal = signal.astype(np.int16).data
data = kaldi.matrix.SubMatrix(
kaldi.feat.pitch.compute_kaldi_pitch(
self._options, kaldi.matrix.SubVector(signal))).numpy()
return Features(
data, self.times(data.shape[0]), properties=self.get_properties())
def test_copy(mfcc):
# by copy we allocate new arrays
mfcc2 = mfcc.copy()
assert mfcc2 == mfcc
assert mfcc2 is not mfcc
assert mfcc2.data is not mfcc.data
assert mfcc2.times is not mfcc.times
assert mfcc2.properties is not mfcc.properties
# by explicit construction the arrays are shared
mfcc2 = Features(mfcc.data,
mfcc.times,
properties=mfcc.properties,
validate=False)
assert mfcc2 == mfcc
assert mfcc2 is not mfcc
assert mfcc2.data is mfcc.data
assert mfcc2.times is mfcc.times
assert mfcc2.properties is mfcc.properties
def process(self, signal):
# ensure the signal is correct
if signal.nchannels != 1:
raise ValueError(
'signal must have one dimension, but it has {}'.format(
signal.nchannels))
if self.sample_rate != signal.sample_rate:
raise ValueError('processor and signal mismatch in sample rates: '
'{} != {}'.format(self.sample_rate,
signal.sample_rate))
# force the signal to be int16
signal = signal.astype(np.int16)
# extract the features
data = self._rastaplp(signal)
return Features(data.T.astype(np.float32),
self.times(data.T.shape[0]),
properties=self.get_properties())
def process(self, features):
"""Compute deltas on `features` with the specified options
Parameters
----------
features : Features, shape = [nframes, ncols]
The input features on which to compute the deltas
Returns
-------
deltas : Features, shape = [nframes, ncols * (`order` + 1)]
The computed deltas with as much orders as specified. The
output features are the concatenation of the input
`features` and it's time derivative at each orders.
"""
data = kaldi.matrix.SubMatrix(
kaldi.feat.functions.compute_deltas(
self._options, kaldi.matrix.SubMatrix(features.data))).numpy()
return Features(data, features.times, self.get_properties(features))
def process(self, alignment):
# build a bijection token <-> onehot index
token2index = self._token2index(alignment)
# initialize the data matrix with zeros, TODO should data be a
# scipy.sparse matrix?
data = np.zeros((alignment.tokens.shape[0], len(token2index)),
dtype=np.bool)
# fill the data with onehot encoding of tokens
for i, p in enumerate(alignment.tokens):
data[i, token2index[p]] = 1
try:
properties = self.get_properties()
except ValueError: # tokens not defined
self.tokens = token2index.keys()
properties = self.get_properties()
self.tokens = None
properties[self.name].update({'token2index': token2index})
return Features(data, alignment.times, properties=properties)
def process(self, signal):
"""Computes energy on the input `signal`
Parameters
----------
signal : :class:`~signal.audio.audioData`
Returns
-------
energy : :class:`~shennong.features.features.Features`
The computed - and compressed - energy
Raises
------
ValueError
If the input `signal` has more than one channel (i.e. is
not mono). If `sample_rate` != `signal.sample_rate`.
"""
# ensure the signal is correct
if signal.nchannels != 1:
raise ValueError(
'signal must have one dimension, but it has {}'.format(
signal.nchannels))
if self.sample_rate != signal.sample_rate:
raise ValueError('processor and signal mismatch in sample rates: '
'{} != {}'.format(self.sample_rate,
signal.sample_rate))
if self.raw_energy:
old_conf = self.get_params()
self.preemph_coeff = 0
self.window_type = 'rectangular'
# number of frames in the framed signal
nframes = kaldi.feat.window.num_frames(signal.nsamples,
self._frame_options,
flush=True)
# a kaldi view of the numpy signal
signal = kaldi.matrix.SubVector(signal.data)
# windowing function to compute frames
window = kaldi.feat.window.FeatureWindowFunction.from_options(
self._frame_options)
# compression function to compress energy
compression = self._compression_fun[self._compression]
# pre-allocate the resulting energy
energy = np.zeros((nframes, 1))
# pre-allocate a buffer for the frames, extract the frames and
# compute the energy on them
out_frame = kaldi.matrix.Vector(self._frame_options.window_size())
for frame in range(nframes):
kaldi.feat.window.extract_window(0, signal, frame,
self._frame_options, window,
out_frame)
# square the signal, force float64 to avoid overflow
square = np.square(out_frame.numpy(), dtype=np.float64)
# avoid doing log on 0 (should be avoided already by
# dithering, but who knows...)
energy[frame] = compression(
max(square.sum(),
np.finfo(np.float64).tiny))
if self.raw_energy:
self.set_params(**old_conf)
return Features(energy, self.times(nframes), self.get_properties())
def process(self, signal):
"""Computes bottleneck features on an audio `signal`
Use a pre-trained neural network to extract bottleneck
features. Features have a frame shift of 10 ms and frame
length of 25 ms.
Parameters
----------
signal : Audio, shape = [nsamples, 1]
The input audio signal to compute the features on, must be
mono. The signal is up/down-sampled at 8 kHz during
processing.
Returns
-------
features : Features, shape = [nframes, 80]
The computes bottleneck features will have as many rows as
there are frames (depends on the `signal` duration, expect
about 100 frames per second), each frame with 80
dimensions.
Raises
------
RuntimeError
If no speech is detected on the `signal` during the voice
activity detection preprocessing step.
"""
# force resampling to 8 kHz and 16 bits integers
need_resample = (signal.sample_rate != 8000
or signal.dtype is not np.dtype(np.int16))
if need_resample:
self._log.debug('resampling audio from %dHz@%db to %dHz@%db',
signal.sample_rate, signal.dtype.itemsize * 8,
8000, 16)
signal = signal.resample(8000).astype(np.int16)
signal = signal.data
# define parameters to extract mel filterbanks. Those
# parameters cannot be tuned because the networks are trained
# with them... frame_noverlap is the number of samples to
# overlap in each frame, so the frame_shift is 200 - 120 = 80
frame_length = 200
frame_noverlap = 120
frame_shift = frame_length - frame_noverlap
# voice activity detection TODO implement user-provided VAD
# (vad input format could be an instance of Alignment, or
# simply an array of bool).
vad = _compute_vad(signal,
self._log,
win_length=frame_length,
win_overlap=frame_noverlap)
# ensure we have some voiced frames in the signal
voiced_frames = sum(vad)
if not voiced_frames:
raise RuntimeError(
'no voice detected in signal, failed to extract features')
self._log.debug('%d frames of speech detected (on %d total frames)',
voiced_frames, len(vad))
# from audio signal to mel filterbank
signal = _add_dither(signal, self.dither)
window = np.hamming(frame_length)
fbank_mx = _mel_fbank_mx(window.size,
8000,
numchans=24,
lofreq=64.0,
hifreq=3800.0)
fea = _fbank_htk(signal, window, frame_noverlap, fbank_mx)
# center the mel features from voiced frames mean
fea -= np.mean(fea[vad], axis=0)
# add a global context to the mel features
left_ctx = right_ctx = 15
fea = np.r_[np.repeat(fea[[0]], left_ctx, axis=0), fea,
np.repeat(fea[[-1]], right_ctx, axis=0)]
# compute the network output from mel features
left_ctx_bn1 = right_ctx_bn1 = self._get_weights()['context']
nn_input = _preprocess_nn_input(fea, left_ctx_bn1, right_ctx_bn1)
nn_output = np.vstack(
_create_nn_extract_st_BN(nn_input, self._get_weights(), 2)[0])
# compute the timestamps for each output frame
times = (1.0 / 8000) * np.vstack(
(np.arange(nn_output.shape[0]) * frame_shift,
np.arange(nn_output.shape[0]) * frame_shift + frame_length)).T
# return the final bottleneck features
return Features(nn_output, times, self.get_properties())
def test_2d_times_unsorted():
with pytest.raises(ValueError) as err:
Features(np.random.random((10, 3)), np.random.random((10, 2)))
assert 'times is not sorted in increasing order' in str(err)
def test_collection(mfcc):
assert FeaturesCollection._value_type is Features
assert FeaturesCollection().is_valid()
assert FeaturesCollection(mfcc=mfcc).is_valid()
assert not FeaturesCollection(
mfcc=Features(np.asarray([0]), 0, validate=False)).is_valid()
def test_2d_times_badshape():
with pytest.raises(ValueError) as err:
Features(np.random.random((10, 3)), np.random.random((10, 3)))
assert 'times shape[1] must be 2, it is 3' in str(err)
