def test_frame_sig(self):
n = 10000124
frame_len = 37
frame_step = 13
x = np.random.rand(n)
t0 = time.time()
y_old = sigproc.framesig(x, frame_len=frame_len, frame_step=frame_step, stride_trick=False)
t1 = time.time()
y_new = sigproc.framesig(x, frame_len=frame_len, frame_step=frame_step, stride_trick=True)
t_new = time.time() - t1
t_old = t1 - t0
self.assertTupleEqual(y_old.shape, y_new.shape)
np.testing.assert_array_equal(y_old, y_new)
self.assertLess(t_new, t_old)
print('new run time %3.2f < %3.2f sec' % (t_new, t_old))
def specdecomp(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfft=512,
lowfreq=0,
highfreq=None,
preemph=0.97,
winfunc=lambda x: np.ones((x, )),
decomp='complex'):
"""Compute the magnitude spectrum of each frame in frames. If frames is an NxD matrix, output will be Nx(NFFT/2+1).
:param frames: the array of frames. Each row is a frame.
:param NFFT: the FFT length to use. If NFFT > frame_len, the frames are zero-padded.
:returns: If frames is an NxD matrix, output will be Nx(NFFT/2+1). Each row will be the magnitude spectrum of the corresponding frame.
"""
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
if decomp == 'time' or decomp == 'frames':
return frames
complex_spec = np.fft.rfft(frames, nfft)
if decomp == 'magnitude' or decomp == 'mag' or decomp == 'abs':
return np.abs(complex_spec)
elif decomp == 'phase' or decomp == 'angle':
return np.angle(complex_spec)
elif decomp == 'power' or decomp == 'powspec':
return sigproc.powspec(frames, nfft)
else:
return complex_spec
return spect
def specspreadcent_xtr_func(sig, fs, args, winlen, winstep):
chnkd_sig = framesig(sig, winlen * fs, winstep * fs)
"""Spectral spread and centroid of windows - based on pyAudioAnalysis.audioFeatureExtraction library
[github.com/tyiannak/pyAudioAnalysis]"""
#chnkd_sigs = raw_chnkd_xtr_func(chnkd_sig_lst, fs_lst, args)
centroids, spreads = [], []
for chnk in chnkd_sig:
spec = get_win_fft(chnk, winlen, fs)
ind = (np.arange(1, len(spec) + 1)) * (fs / (2.0 * len(spec)))
Xt = spec.copy()
Xt = Xt / Xt.max()
NUM = np.sum(ind * Xt)
DEN = np.sum(Xt) + eps
# Centroid:
C = (NUM / DEN)
# Spread:
S = np.sqrt(np.sum(((ind - C)**2) * Xt) / DEN)
# Normalize:
C = C / (fs / 2.0)
S = S / (fs / 2.0)
centroids.append(C)
spreads.append(S)
res = [[cent, spread] for cent, spread in zip(centroids, spreads)]
return np.array(res)
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97,
winfunc=lambda x:numpy.ones((x,))):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied.
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate, winfunc)
pspec = sigproc.powspec(frames,nfft)
energy = numpy.sum(pspec,1) # this stores the total energy in each frame
energy = numpy.where(energy == 0,numpy.finfo(float).eps,energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
feat = numpy.where(feat == 0,numpy.finfo(float).eps,feat) # if feat is zero, we get problems with log
return feat,energy
def ssc(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97,
winfunc=lambda x:numpy.ones((x,))):
"""Compute Spectral Subband Centroid features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied.
:returns: A numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector.
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate, winfunc)
pspec = sigproc.powspec(frames,nfft)
pspec = numpy.where(pspec == 0,numpy.finfo(float).eps,pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
R = numpy.tile(numpy.linspace(1,samplerate/2,numpy.size(pspec,1)),(numpy.size(pspec,0),1))
return numpy.dot(pspec*R,fb.T) / feat
def compute_rmfcc(self, pad, save=False):
filepaths = self.wavpaths
frames = {}
fno = 0
self.af = {}
self.rmfcc = {}
for f in filepaths:
if fno%10 == 0:
print fno
fno += 1
sig, sr = librosa.load(f[0])
frames = sigproc.framesig(sig, frame_len=500, frame_step=250)
af = []
rmfcc = []
for frame in frames:
analysis_filt = lpc.autocor(frame, 12)
af.append(analysis_filt)
residual = np.array(list(analysis_filt(frame)))
temp = list(python_speech_features.mfcc(frame, sr, winlen=0.022))
rmfcc.append(temp[0])
rmfcc,_ = self.pad_sequence_into_array(np.array(rmfcc).transpose(), maxlen=pad)
print rmfcc.shape
self.rmfcc[f[1]] = rmfcc
self.af[f[1]] = af
with open('analysis_filt.pkl', 'wb') as f:
pickle.dump(self.af, f)
def extract_features_file(filename):
"""Extract feature vectors for file
Frames the file according to WINLEN and WINSTEP and extracts a featurevector for every frame.
:param filename: filename of the IRMAS dataset
:return: 2D-numpy array with the shape (numframes, numfeatures) NOTE: if numframes == 1 it returns a single featurevector
"""
# read wav file
data = wav.read(filename)
data = data[1]
data = data[:, 1]
frames = framesig(data, winlen_samp, winlen_samp)
features = None
for frame in frames:
if features is None:
features = extract_features_window(frame)
else:
new_features = extract_features_window(frame)
features = np.vstack((features, new_features))
return features
def ssc(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfilt=55,
nfft=2048,
lowfreq=0,
highfreq=None,
preemph=0.97,
winfunc=lambda x: numpy.ones((x, ))):
"""Compute Spectral Subband Centroid features from an audio signal.
"""
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
pspec = sigproc.powspec(frames, nfft)
pspec = numpy.where(pspec == 0,
numpy.finfo(float).eps,
pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt, nfft, samplerate, lowfreq, highfreq)
feat = numpy.dot(pspec, fb.T) # compute the filterbank energies
R = numpy.tile(numpy.linspace(1, samplerate / 2, numpy.size(pspec, 1)),
(numpy.size(pspec, 0), 1))
return numpy.dot(pspec * R, fb.T) / feat
def fbank(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfilt=55,
nfft=2048,
lowfreq=0,
highfreq=None,
preemph=0.97,
winfunc=lambda x: numpy.ones((x, ))):
"""Compute Mel-filterbank energy features from an audio signal.
"""
highfreq = highfreq or samplerate / 2
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
pspec = sigproc.powspec(frames, nfft)
energy = numpy.sum(pspec, 1) # this stores the total energy in each frame
energy = numpy.where(energy == 0,
numpy.finfo(float).eps,
energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt, nfft, samplerate, lowfreq, highfreq)
feat = numpy.dot(pspec, fb.T) # compute the filterbank energies
feat = numpy.where(feat == 0,
numpy.finfo(float).eps,
feat) # if feat is zero, we get problems with log
return feat, energy
def cspec(signal,
samplerate=16000,
winlen=0.025,
winstep=0.01,
nfft=512,
preemph=0.97,
winfunc=lambda x: np.ones((x, ))):
"""Compute STFT coeeficients from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfft: the FFT size. Default is 512.
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied.
You can use numpy window functions here e.g. winfunc=np.hamming
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features.
Each row holds 1 feature vector. The second return value is the energy in each frame (total energy, unwindowed)
"""
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
if np.shape(frames)[1] > nfft:
logging.warn(
'frame length (%d) is greater than FFT size (%d), frame will be truncated. '
+ 'Increase NFFT to avoid.',
np.shape(frames)[1], nfft)
return np.fft.rfft(frames, nfft)
def fbank(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97,
winfunc=lambda x:numpy.ones((x,))):
"""Compute Mel-filterbank energy features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied. You can use numpy window functions here e.g. winfunc=numpy.hamming
:returns: 2 values. The first is a numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector. The
second return value is the energy in each frame (total energy, unwindowed)
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate, winfunc)
pspec = sigproc.powspec(frames,nfft)
energy = numpy.sum(pspec,1) # this stores the total energy in each frame
energy = numpy.where(energy == 0,numpy.finfo(float).eps,energy) # if energy is zero, we get problems with log
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
feat = numpy.where(feat == 0,numpy.finfo(float).eps,feat) # if feat is zero, we get problems with log
return feat,energy
def ssc(signal,samplerate=16000,winlen=0.025,winstep=0.01,
nfilt=26,nfft=512,lowfreq=0,highfreq=None,preemph=0.97,
winfunc=lambda x:numpy.ones((x,))):
"""Compute Spectral Subband Centroid features from an audio signal.
:param signal: the audio signal from which to compute features. Should be an N*1 array
:param samplerate: the samplerate of the signal we are working with.
:param winlen: the length of the analysis window in seconds. Default is 0.025s (25 milliseconds)
:param winstep: the step between successive windows in seconds. Default is 0.01s (10 milliseconds)
:param nfilt: the number of filters in the filterbank, default 26.
:param nfft: the FFT size. Default is 512.
:param lowfreq: lowest band edge of mel filters. In Hz, default is 0.
:param highfreq: highest band edge of mel filters. In Hz, default is samplerate/2
:param preemph: apply preemphasis filter with preemph as coefficient. 0 is no filter. Default is 0.97.
:param winfunc: the analysis window to apply to each frame. By default no window is applied. You can use numpy window functions here e.g. winfunc=numpy.hamming
:returns: A numpy array of size (NUMFRAMES by nfilt) containing features. Each row holds 1 feature vector.
"""
highfreq= highfreq or samplerate/2
signal = sigproc.preemphasis(signal,preemph)
frames = sigproc.framesig(signal, winlen*samplerate, winstep*samplerate, winfunc)
pspec = sigproc.powspec(frames,nfft)
pspec = numpy.where(pspec == 0,numpy.finfo(float).eps,pspec) # if things are all zeros we get problems
fb = get_filterbanks(nfilt,nfft,samplerate,lowfreq,highfreq)
feat = numpy.dot(pspec,fb.T) # compute the filterbank energies
R = numpy.tile(numpy.linspace(1,samplerate/2,numpy.size(pspec,1)),(numpy.size(pspec,0),1))
return numpy.dot(pspec*R,fb.T) / feat
def stft(sig, rate):
frames = sigproc.framesig(sig,
FRAME_LENGTH*rate,
FRAME_SHIFT*rate,
winfunc=squared_hann)
spec = np.fft.rfft(frames, int(FRAME_LENGTH*rate))
# adding 1e-7 just to avoid problems with log(0)
return np.log10(np.absolute(spec)+1e-7) # Log 10 for easier dB calculation
def prepare_tfrecord(example_paths,
destination_path,
max_len,
nfft=256,
noverlap=128):
'''
Converts a set of inputs into spectrograms and saves them to disk in
a TensorFlow native format.
:return A tuple container (min_val, max_val. mean) for the set.
'''
features_min, features_max = None, None
features_count, features_sum = 0, 0
# Open a TFRecords file for writing.
writer = tf.python_io.TFRecordWriter(destination_path)
for idx in range(len(example_paths)):
# Load an audio file for preprocessing.
try:
samples, _ = librosa.load(example_paths[idx])
except NoBackendError:
print('Warning: Could not load {}.'.format(example_paths[idx]))
continue
# Pad or shorten the number audio samples to max length.
if samples.shape[0] < max_len:
samples = np.pad(samples, (0, max_len - samples.shape[0]),
'constant')
elif samples.shape[0] > max_len:
samples = samples[:max_len]
# Generate a log power spectrum of the audio samples.
spectrum = np.abs(
logpowspec(framesig(samples, nfft, noverlap, winfunc=np.hanning),
nfft,
norm=0))
spectrum = np.transpose(np.flip(spectrum, 1)).astype(np.float32)
label = int(os.path.split(example_paths[idx])[-1].split('-')[1])
# Keep track of the dataset statistics.
new_min = np.min(spectrum)
new_max = np.max(spectrum)
if features_min is not None and features_max is not None:
features_min = new_min if features_min > new_min else features_min
features_max = new_max if features_max < new_max else features_max
else:
features_min = new_min
features_max = new_max
features_count += np.prod(spectrum.shape)
features_sum += np.sum(spectrum)
# Write the final spectrum and label to disk.
example = tf.train.Example(features=tf.train.Features(
feature={
'spectrum': bytes_feature(spectrum.flatten().tostring()),
'label': int64_feature(label)
}))
writer.write(example.SerializeToString())
writer.close()
# Return the dataset statistics.
return features_min, features_max, float(features_sum) / features_count
def load_frames(in_file, srate=16000):
'''load frames from either single wav or npy
@in_file: input wav or npy
@return: raw frames in the size of [num_frames, frame_len]
'''
signal, srate = librosa.load(in_file, srate)
pre_emphed = sigproc.preemphasis(signal, coeff=0.95)
frames = sigproc.framesig(pre_emphed, 0.025 * srate, 0.01 * srate)
return frames
def calc_sp(mix, clean, data_type, Win_Length, Offset_Length):
"""
This func is to calculate the features and corresponding labels in the time domain
:param mix: 1-D vector
:param clean: 1-D vector
:param data_type:
:param Win_Length: the length of the window function
:param Offset_Length: the offset length between adjanct frames
:return:
""" """
"""
n_window = Win_Length
n_offset = Offset_Length
mix_x = framesig(mix, frame_len=n_window, frame_step=n_offset)
clean_x = framesig(clean, frame_len=n_window, frame_step=n_offset)
return mix_x, clean_x
def zcr_xtr_func(sig, fs, args, winlen, winstep):
chnkd_sig = framesig(sig, winlen * fs, winstep * fs)
"""Sign-change rate of signal per frame."""
#chnkd_sig = raw_chnkd_xtr_func(chnkd_sig_lst, fs, args)
zcr_wins = []
for chnk in chnkd_sig:
zcr_win = np.sum(chnk[:-1] * chnk[1:] < 0)
zcr_wins.append(zcr_win)
return np.array(zcr_wins)
def _powspec_cord(self, data, fs):
kwargs = self._cording_params
preemph = kwargs["preemph"]
winlen = kwargs["winlen"]
winstep = kwargs["winstep"]
winfunc = kwargs["winfunc"]
fft_size = int(winlen * fs)
data = preemphasis(data, preemph)
frames = framesig(data, winlen * fs, winstep * fs, winfunc)
return powspec(frames, fft_size)
def extract(sig):
# framing
sig_frames = sigproc.framesig(sig=sig, frame_len=FRAME_LENGTH, frame_step=FRAME_STEP)
frames_feats = None
def concat_feats(feat_coeffs):
return np.concatenate((frames_feats, feat_coeffs), axis=1)
# region calculate mfcc features
mfcc_feat = mfcc(signal=sig, samplerate=SAMPLE_RATE, winlen=WINDOW_LENGTH, winstep=WINDOW_STEP,
numcep=13, preemph=PRE_EMPH, winfunc=WINDOW_FUNCTION)
mfcc_feat_delta = delta(mfcc_feat, 20)
mfcc_feat_delta_delta = delta(mfcc_feat_delta, 20)
frames_feats = mfcc_feat
frames_feats = concat_feats(mfcc_feat_delta)
frames_feats = concat_feats(mfcc_feat_delta_delta)
# endregion
# region calculate zero cross rating
def zcr(frames):
def sign(x):
return 1 if x >= 0 else -1
zcrs = []
for frame in frames:
zc_rate = 0
for i in range(1, len(frame)):
zc_rate += abs(sign(frame[i]) - sign(frame[i - 1])) / 2
zcrs.append(zc_rate / len(frame))
return zcrs
zcrs = zcr(sig_frames)
frames_feats = concat_feats(np.array([zcrs]).reshape(len(zcrs), 1))
# endregion
# region calculate energy
def autocorrelate(frames, eta):
energys = []
for frame in frames:
total_sum = 0
for i in range(eta, len(frame)):
total_sum += frame[i] * frame[i - eta]
energy = 1 / len(frame) * total_sum
energys.append(energy)
return energys
energys = autocorrelate(sig_frames, 0)
frames_feats = concat_feats(np.array([energys]).reshape(len(energys), 1))
# endregion
# frames_feats = frames_feats/frames_feats.max(axis=1).reshape(frames_feats.shape[0],1)
return frames_feats/100000
def parse_audio(self, audio_path):
if self.augment:
y = load_randomly_augmented_audio(audio_path, self.sample_rate)
else:
y = load_audio(audio_path)
if self.noiseInjector:
add_noise = np.random.binomial(1, self.noise_prob)
if add_noise:
y = self.noiseInjector.inject_noise(y)
# Split audio into frames
frame_len_ = self.sample_rate*self.window_size
frame_step_ = self.sample_rate*self.window_stride
frames = sigproc.framesig(y,frame_len=frame_len_,frame_step=frame_step_)
# Compute features
features = None
if self.feature_type=='rawspeech':
# Raw speech signal (dimension = 1 X # of samples)
y = y.reshape((1,len(y)))
features = y
features = torch.FloatTensor(features)
elif self.feature_type=='rawframes':
# Raw speech frames (dimension = # of frames X frame length)
features = frames
features = torch.FloatTensor(features.transpose())
elif self.feature_type=='spectrogram':
# Spectrogram
features = sigproc.magspec(frames,NFFT=int(frame_len_))
features = torch.FloatTensor(features.transpose())
elif self.feature_type=='mfcc':
# MFCCs
mfcc_feat = mfcc(y,self.sample_rate,winlen=self.window_size,winstep=self.window_stride,
numcep=13,nfilt=26)
delta = mfccdelta(mfcc_feat,2)
deltadelta = mfccdelta(delta,2)
mfcc_feat = torch.FloatTensor(mfcc_feat.transpose())
delta = torch.FloatTensor(delta.transpose())
deltadelta = torch.FloatTensor(deltadelta.transpose())
features = torch.cat((mfcc_feat,delta,deltadelta),0)
elif self.feature_type=='logmel':
# Log Mel-FB features
logmel_feat = logfbank(y,self.sample_rate,winlen=self.window_size,winstep=self.window_stride,nfilt=26)
delta = mfccdelta(logmel_feat,2)
deltadelta = mfccdelta(delta,2)
logmel_feat = torch.FloatTensor(logmel_feat.transpose())
delta = torch.FloatTensor(delta.transpose())
deltadelta = torch.FloatTensor(deltadelta.transpose())
features = torch.cat((logmel_feat,delta,deltadelta),0)
if self.normalize:
mean = torch.mean(features,0,keepdim=True)
mean = torch.cat([mean]*features.size(0))
std = torch.std(features,0,keepdim=True)
std =torch.cat([std]*features.size(0))
features = (features-mean)/std
return features
def energy_xtr_func(sig, fs, args, winlen, winstep):
chnkd_sig = framesig(sig, winlen * fs, winstep * fs)
"""Sum of squares of signal values, normalised by window length."""
#chnkd_sigs = raw_chnkd_xtr_func(chnkd_sig_lst, fs_lst, args)
#for chnkd_sig, fs in zip(chnkd_sigs, fs_lst):
energy_wins = []
for chnk in chnkd_sig:
nrm_energy_win = 1. / len(chnk) * np.sum(chnk**2)
energy_wins.append(nrm_energy_win)
return np.array(energy_wins)
def power_(signal,
samplerate=16000,
winlen=0.08,
winstep=0.04,
nfft=2048,
preemph=0.97,
winfunc=lambda x: numpy.ones((x, ))):
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
feat = sigproc.powspec(frames, nfft)
return feat
def get_features(file_name):
(rate, sig) = wavfile.read(file_name)
with contextlib.closing(wave.open(file_name, 'r')) as f:
frames = f.getnframes()
rate = f.getframerate()
duration = frames / float(rate)
# Frame our signal into 20 frames with 50% overlap
number_of_frames = 40
frame_len = len(sig) / (number_of_frames * (.5) + .5)
frames = framesig(sig, frame_len, frame_len * .5)
# A list of 20 frequency lists for each frame. 6 frequency bands with the average energy of each
features = []
band0 = []
band1 = []
band2 = []
band3 = []
band4 = []
band5 = []
for frame in frames:
spectrum, freqs, t, img = specgram(frame, Fs=rate)
i = 0
bands = []
for freq in freqs:
if freq <= 400:
band0.extend(spectrum[i])
elif freq > 400 and freq <= 800:
band1.extend(spectrum[i])
elif freq > 800 and freq <= 1600:
band2.extend(spectrum[i])
elif freq > 1600 and freq <= 2800:
band3.extend(spectrum[i])
elif freq > 2800 and freq <= 4400:
band4.extend(spectrum[i])
elif freq > 4400:
band5.extend(spectrum[i])
i += 1
bands.append(sum(band0) / len(band0))
bands.append(sum(band1) / len(band1))
bands.append(sum(band2) / len(band2))
bands.append(sum(band3) / len(band3))
bands.append(sum(band4) / len(band4))
bands.append(sum(band5) / len(band5))
features.append(bands)
values = []
for feature in features:
for f in feature:
values.append(f)
return values
def Make_Spect(wav_path,
windowsize,
stride,
window=np.hamming,
bandpass=False,
lowfreq=0,
highfreq=0,
preemph=0.97,
duration=False,
nfft=None,
normalize=True):
"""
read wav as float type. [-1.0 ,1.0]
:param wav_path:
:param windowsize:
:param stride:
:param window: default to np.hamming
:return: return spectrogram with shape of (len(wav/stride), windowsize * samplerate /2 +1).
"""
# samplerate, samples = wavfile.read(wav_path)
samples, samplerate = sf.read(wav_path, dtype='float32')
if bandpass and highfreq > lowfreq:
samples = butter_bandpass_filter(data=samples,
cutoff=[lowfreq, highfreq],
fs=samplerate)
signal = sigproc.preemphasis(samples, preemph)
frames = sigproc.framesig(signal,
windowsize * samplerate,
stride * samplerate,
winfunc=window)
if nfft == None:
nfft = int(windowsize * samplerate)
pspec = sigproc.powspec(frames, nfft)
pspec = np.where(pspec == 0, np.finfo(float).eps, pspec)
# S = librosa.stft(samples, n_fft=int(windowsize * samplerate),
# hop_length=int((windowsize-stride) * samplerate),
# window=window(int(windowsize * samplerate))) # 进行短时傅里叶变换，参数意义在一开始有定义
# feature, _ = librosa.magphase(S)
# feature = np.log1p(feature) # log1p操作
feature = np.log(pspec).astype(np.float32)
# feature = feature.transpose()
if normalize:
feature = normalize_frames(feature)
if duration:
return feature, len(samples) / samplerate
return feature
def test_frame_sig(self):
n = 10000124
frame_len = 37
frame_step = 13
x = np.random.rand(n)
t0 = time.time()
y_old = sigproc.framesig(x,
frame_len=frame_len,
frame_step=frame_step,
stride_trick=False)
t1 = time.time()
y_new = sigproc.framesig(x,
frame_len=frame_len,
frame_step=frame_step,
stride_trick=True)
t_new = time.time() - t1
t_old = t1 - t0
self.assertTupleEqual(y_old.shape, y_new.shape)
np.testing.assert_array_equal(y_old, y_new)
self.assertLess(t_new, t_old)
print('new run time %3.2f < %3.2f sec' % (t_new, t_old))
def datafeed():
data_w.put('start')
loc, sec, i = 0, 0, 0
if '.wav' not in args.track:
rsmpfile = 'resampled.wav'
resampling(args.track, rsmpfile)
else:
rsmpfile = args.track
data_wav, fs = soundfile.read(rsmpfile)
if fs != 16000:
rsmpfile = 'resampled.wav'
resampling(args.track, rsmpfile)
data_wav, fs = soundfile.read(rsmpfile)
data_wav /= np.amax(np.abs(data_wav))
idxs = np.linspace(0, fs, 31, endpoint=True, dtype=np.int)
rest = [0.0325, 0.0335, 0.0325]
slope = (rng[1] - rng[0]) / (audio_max - audio_min)
intersec = rng[1] - slope * audio_max
data_w.put('start')
if vlclib:
vlcplayer.play()
if enable_record:
ws.call(requests.StartRecording())
NFFT = int(2**(np.ceil(np.log2(256))))
while loc < data_wav.shape[0]:
# t = time.time()
prv = idxs[i] + fs * sec
loc = idxs[i + 1] + fs * sec
frames = framesig(data_wav[prv:loc],
256,
80,
winfunc=lambda x: np.hamming(x))
stft_data = logpowspec(frames, NFFT) # , fs, 160, 80)
stft_data = (stft_data * slope) + intersec
stft_data = np.swapaxes(stft_data, 0, 1).astype(np.float32)
if stft_data.shape[1] != 5:
stft_data = np.ones((129, 5), dtype=np.float32) * rng[0]
data_w.put((0, stft_data.copy()))
break
data_w.put((0, stft_data.copy()))
sleep(rest[i % 3])
if i >= 29:
i = 0
sec += 1
else:
i += 1
os.remove(rsmpfile)
data_w.put('end')
return
def fft_sam(signal,
samplerate=16000,
winlen=0.08,
winstep=0.04,
nfft=2048,
preemph=0.97,
winfunc=lambda x: numpy.ones((x, ))):
signal = sigproc.preemphasis(signal, preemph)
frames = sigproc.framesig(signal, winlen * samplerate,
winstep * samplerate, winfunc)
feat = numpy.float32(numpy.absolute(numpy.fft.fft(
frames, nfft)))[:, 0:int(numpy.floor(nfft / 2)) + 1]
for i in range(0, len(feat)):
feat[i, 1:] = feat[i, 1:] - feat[i, :-1]
return feat
def get_complex_spec(wav_, winstep, winlen, with_time_scaled=False):
"""Return complex spec
"""
rate, sig = wav.read(wav_)
sig = preemphasis(sig, PREEMPH)
frames = framesig(sig, winlen * rate, winstep * rate, HAMMING_WINFUNC)
complex_spec = np.fft.rfft(frames, NFFT)
time_scaled_complex_spec = None
if with_time_scaled:
time_scaled_frames = np.arange(frames.shape[-1]) * frames
time_scaled_complex_spec = np.fft.rfft(time_scaled_frames, NFFT)
return complex_spec, time_scaled_complex_spec
def job(input_name, output_name):
audio, _ = librosa.load(input_name, mono=True, sr=samplerate)
if len(audio) == 0:
return False
signal = sigproc.preemphasis(audio, 0.97)
x = sigproc.framesig(signal, winlen, winstep, np.hanning)
if len(x) == 0:
return False
x = sigproc.powspec(x, nfft)
x = np.dot(x, banks)
x = np.where(x == 0, np.finfo(float).eps, x)
x = np.log(x).astype(dtype=np.float32)
if np.isnan(np.sum(x)):
return False
np.save(output_name, x)
return True
def energyentropy_xtr_func(sig, fs, args, winlen, winstep):
chnkd_sig = framesig(sig, winlen * fs, winstep * fs)
"""Entropy of energy - based on pyAudioAnalysis.audioFeatureExtraction library
[github.com/tyiannak/pyAudioAnalysis]"""
#chnkd_sigs = raw_chnkd_xtr_func(chnkd_sig_lst, fs_lst, args)
entropies = []
for chnk in chnkd_sig:
tot_enrgy = np.sum(chnk**2)
subwin_len = int(np.floor(len(chnk) / args["n_subwins"]))
if len(chnk) != subwin_len * args["n_subwins"]:
chnk = chnk[0:subwin_len * args["n_subwins"]]
subwins = chnk.reshape(subwin_len, args["n_subwins"], order='F').copy()
subwin_enrgy = np.sum(subwins**2, axis=0) / float(tot_enrgy + eps)
entropy = -np.sum(subwin_enrgy * np.log2(subwin_enrgy + eps))
entropies.append(entropy)
return np.array(entropies)
def compute_lpcc(self, pad, save=False):
filepaths = self.wavpaths
self.lpcc = {}
self.sr = {}
self.frames = {}
st = time.time()
for f in filepaths:
sig, self.sr[f[1]] = librosa.load(f[0])
self.frames[f[1]] = sigproc.framesig(sig, frame_len=2200, frame_step=1100)
temp = list(self.frames[f[1]])
arr = np.array([np.array(lpc.kautocor(i, 12).numerator) for i in temp])
print arr.shape
self.lpcc[f[1]] = self.pad_sequence_into_array(arr.transpose(), maxlen=pad)
if save:
with open('feat/lpcc.pkl', 'wb') as f:
pickle.dump(self.lpcc, f)
def specflux_xtr_func(sig, fs, args, winlen, winstep):
chnkd_sig = framesig(sig, winlen * fs, winstep * fs)
"""Spectral flux as sum of square differences - based on pyAudioAnalysis.audioFeatureExtraction library
[github.com/tyiannak/pyAudioAnalysis]"""
#chnkd_sigs = raw_chnkd_xtr_func(chnkd_sig_lst, fs_lst, args)
fluxs, prev_chnk, prev_chnk_sum = [], [], []
for chnk in chnkd_sig:
spec = get_win_fft(chnk, winlen, fs)
specsum = np.sum(spec + eps)
if prev_chnk != []:
flux = np.sum((spec / specsum - prev_chnk / prev_chnk_sum)**2)
fluxs.append(flux)
else:
fluxs.append(0.)
prev_chnk = spec
prev_chnk_sum = specsum
return np.array(fluxs)
def get_complex_spec(wav_, winstep, winlen, with_time_scaled=False):
"""Return complex spec
"""
sig, rate = librosa.load(wav_, sr=sr)
#print(rate,sig)
sig = preemphasis(sig, PREEMPH)
frames = framesig(sig, winlen * rate, winstep * rate, HAMMING_WINFUNC)
complex_spec = np.fft.rfft(frames, NFFT)
time_scaled_complex_spec = None
if with_time_scaled:
time_scaled_frames = np.arange(frames.shape[-1]) * frames
time_scaled_complex_spec = np.fft.rfft(time_scaled_frames, NFFT)
print(complex_spec.shape, time_scaled_complex_spec.shape)
return complex_spec, time_scaled_complex_spec
