def __test_hybrid_cqt(pad_mode):
D1 = librosa.hybrid_cqt(y, pad_mode='reflect')
D2 = librosa.hybrid_cqt(y, pad_mode=pad_mode)
assert D1.shape == D2.shape
if pad_mode != 'reflect':
assert not np.allclose(D1, D2)
else:
assert np.allclose(D1, D2)
def __test_hybrid_cqt(pad_mode):
D1 = librosa.hybrid_cqt(y, pad_mode='reflect')
D2 = librosa.hybrid_cqt(y, pad_mode=pad_mode)
assert D1.shape == D2.shape
if pad_mode != 'reflect':
assert not np.allclose(D1, D2)
else:
assert np.allclose(D1, D2)
def test_hybrid_cqt_multi(y_multi, scale, res_type):
y, sr = y_multi
# Assuming single-channel CQT is well behaved
C0 = librosa.hybrid_cqt(y=y[0], sr=sr, scale=scale, res_type=res_type)
C1 = librosa.hybrid_cqt(y=y[1], sr=sr, scale=scale, res_type=res_type)
Call = librosa.hybrid_cqt(y=y, sr=sr, scale=scale, res_type=res_type)
# Check each channel
assert np.allclose(C0, Call[0])
assert np.allclose(C1, Call[1])
# Verify that they're not all the same
assert not np.allclose(Call[0], Call[1])
def compute_features(self):
"""Actual implementation of the features.
Returns
-------
pcp: np.array(N, F)
The features, each row representing a feature vector for a give
time frame/beat.
"""
audio_harmonic, _ = self.compute_HPSS()
pcp_cqt = (
np.abs(
librosa.hybrid_cqt(
audio_harmonic,
sr=self.sr,
hop_length=self.hop_length,
n_bins=self.n_bins,
norm=self.norm,
fmin=self.f_min,
)
)
** 2
)
pcp = librosa.feature.chroma_cqt(
C=pcp_cqt, sr=self.sr, hop_length=self.hop_length, n_octaves=self.n_octaves, fmin=self.f_min
).T
return pcp
def __test(hop_length, fmin, n_bins, bins_per_octave,
tuning, resolution, norm, sparsity):
C2 = librosa.hybrid_cqt(y, sr=sr,
hop_length=hop_length,
fmin=fmin, n_bins=n_bins,
bins_per_octave=bins_per_octave,
tuning=tuning, resolution=resolution,
norm=norm,
sparsity=sparsity)
C1 = librosa.cqt(y, sr=sr,
hop_length=hop_length,
fmin=fmin, n_bins=n_bins,
bins_per_octave=bins_per_octave,
tuning=tuning, resolution=resolution,
norm=norm,
sparsity=sparsity)
eq_(C1.shape, C2.shape)
# Check for numerical comparability
idx1 = (C1 > 1e-4 * C1.max())
idx2 = (C2 > 1e-4 * C2.max())
perc = 0.99
thresh = 1e-3
idx = idx1 | idx2
assert np.percentile(np.abs(C1[idx] - C2[idx]),
perc) < thresh * max(C1.max(), C2.max())
def __test(hop_length, fmin, n_bins, bins_per_octave, tuning, resolution, norm, sparsity):
C2 = librosa.hybrid_cqt(
y,
sr=sr,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
bins_per_octave=bins_per_octave,
tuning=tuning,
resolution=resolution,
norm=norm,
sparsity=sparsity,
)
C1 = librosa.cqt(
y,
sr=sr,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
bins_per_octave=bins_per_octave,
tuning=tuning,
resolution=resolution,
norm=norm,
sparsity=sparsity,
)
eq_(C1.shape, C2.shape)
# Check for numerical comparability
assert np.mean(np.abs(C1 - C2)) < 1e-3
def audio_extract_pcp(
audio,
sr,
n_fft=4096,
hop_len=int(4096 * 0.75),
pcp_bins=84,
pcp_norm=np.inf,
pcp_f_min=27.5,
pcp_n_octaves=6):
audio_harmonic, _ = librosa.effects.hpss(audio)
pcp_cqt = np.abs(librosa.hybrid_cqt(
audio_harmonic,
sr=sr,
hop_length=hop_len,
n_bins=pcp_bins,
norm=pcp_norm,
fmin=pcp_f_min)) ** 2
pcp = librosa.feature.chroma_cqt(
C=pcp_cqt,
sr=sr,
hop_length=hop_len,
n_octaves=pcp_n_octaves,
fmin=pcp_f_min).T
return pcp
def test_hybrid_cqt(
y_hybrid,
sr,
hop_length,
fmin,
n_bins,
bins_per_octave,
tuning,
resolution,
norm,
sparsity,
res_type,
):
# This test verifies that hybrid and full cqt agree down to 1e-4
# on 99% of bins which are nonzero (> 1e-8) in either representation.
C2 = librosa.hybrid_cqt(
y_hybrid,
sr=sr,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
bins_per_octave=bins_per_octave,
tuning=tuning,
filter_scale=resolution,
norm=norm,
sparsity=sparsity,
res_type=res_type,
)
C1 = np.abs(
librosa.cqt(
y_hybrid,
sr=sr,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
bins_per_octave=bins_per_octave,
tuning=tuning,
filter_scale=resolution,
norm=norm,
sparsity=sparsity,
res_type=res_type,
))
assert C1.shape == C2.shape
# Check for numerical comparability
idx1 = C1 > 1e-4 * C1.max()
idx2 = C2 > 1e-4 * C2.max()
perc = 0.99
thresh = 1e-3
idx = idx1 | idx2
assert np.percentile(np.abs(C1[idx] - C2[idx]),
perc) < thresh * max(C1.max(), C2.max())
def __test(sr, hop_length, y):
hcqt = librosa.hybrid_cqt(y=y, sr=sr, hop_length=hop_length, tuning=0)
response = np.mean(np.abs(hcqt)**2, axis=1)
continuity = np.abs(np.diff(response))
assert np.max(continuity) < 5e-4, continuity
def __test(sr, hop_length, y):
hcqt = librosa.hybrid_cqt(y=y, sr=sr, hop_length=hop_length, tuning=0)
response = np.mean(np.abs(hcqt)**2, axis=1)
continuity = np.abs(np.diff(response))
assert np.max(continuity) < 5e-4, continuity
def test_hybrid_cqt_white_noise(y_white, sr_white, fmin, n_bins, scale):
C = librosa.hybrid_cqt(y=y_white, sr=sr_white, fmin=fmin, n_bins=n_bins, scale=scale)
if not scale:
lengths = librosa.filters.constant_q_lengths(sr_white, fmin, n_bins=n_bins)
C /= np.sqrt(lengths[:, np.newaxis])
assert np.allclose(np.mean(C, axis=1), 1.0, atol=2.5e-1), np.mean(C, axis=1)
assert np.allclose(np.std(C, axis=1), 0.5, atol=5e-1), np.std(C, axis=1)
def extract_features(self, y, sr):
try:
if self.params['normalize']:
rms = np.sqrt(np.mean(y * y))
if rms > 1e-4:
y = y / rms
except KeyError:
pass
try:
if self.params['remove_silence']:
y = self.remove_silence(y,
window=32,
hop=32,
threshold=self.params['sil_threshold'])
except KeyError:
pass
if self.params['method'] == 'FFT':
x = np.abs(
librosa.stft(y,
n_fft=self.params['n_fft'],
hop_length=self.params['hop_length']))
x = librosa.logamplitude(x**2)
#x = np.abs(librosa.stft(y,n_fft=320))
elif self.params['method'] == 'Mel Spectrogram':
x = librosa.feature.melspectrogram(
y,
sr,
n_fft=self.params['n_fft'],
hop_length=self.params['hop_length'],
n_mels=self.params['n_mels'])
x = librosa.logamplitude(x**2)
#x = librosa.feature.melspectrogram(y,sr,n_fft=320,n_mels=160)
elif self.params['method'] == 'CQT':
x = librosa.hybrid_cqt(
y,
sr,
hop_length=self.params['hop_length'],
n_bins=self.params['n_bins'],
bins_per_octave=self.params['bins_per_octave'])
x = librosa.logamplitude(x**2)
#x = librosa.hybrid_cqt(y,sr, hop=128, n_bins=144, bins_per_octave=24)
elif self.params['method'] == 'MFCC':
x = librosa.feature.mfcc(y,
sr,
n_fft=self.params['n_fft'],
n_mels=self.params['n_mels'],
n_mfcc=self.params['n_mfcc'],
hop_length=self.params['hop_length'])
delta = librosa.feature.delta(x)
d_delta = librosa.feature.delta(x, order=2)
x = np.concatenate([x, delta, d_delta], axis=0)
return torch.FloatTensor(x)
def test_hybrid_cqt_impulse(y_impulse, sr_impulse, hop_impulse):
# Test to resolve issue #341
# Updated in #417 to use integrated energy instead of pointwise max
hcqt = librosa.hybrid_cqt(y=y_impulse, sr=sr_impulse, hop_length=hop_impulse, tuning=0)
response = np.mean(np.abs(hcqt) ** 2, axis=1)
continuity = np.abs(np.diff(response))
assert np.max(continuity) < 5e-4, continuity
def nietoPCP(self, samples: Signal):
sr = samples.sampleRate
hop_length = self.parameters["hopLength"].value
pcp_sr = sr / hop_length
audio_harmonic, _ = librosa.effects.hpss(samples.values)
# I double checked, and the parameters are the one used in MSAF. 7 octave in pcp_cqt and 6 octaves in pcp
pcp_cqt = np.abs(librosa.hybrid_cqt(audio_harmonic, sr=sr, hop_length=hop_length, n_bins=7 * 12, norm=np.inf,
fmin=27.5))**2
pcp = librosa.feature.chroma_cqt(C=pcp_cqt, sr=sr, hop_length=hop_length, n_octaves=6, fmin=27.5).T
return (Signal(pcp, sampleRate=pcp_sr), )
def __test(fmin, n_bins, scale, sr, y):
C = librosa.hybrid_cqt(y=y, sr=sr,
fmin=fmin,
n_bins=n_bins,
scale=scale)
if not scale:
lengths = librosa.filters.constant_q_lengths(sr, fmin,
n_bins=n_bins)
C /= np.sqrt(lengths[:, np.newaxis])
assert np.allclose(np.mean(C, axis=1), 1.0, atol=2.5e-1), np.mean(C, axis=1)
assert np.allclose(np.std(C, axis=1), 0.5, atol=5e-1), np.std(C, axis=1)
def __test(sr, hop_length, y):
hcqt = librosa.hybrid_cqt(y=y, sr=sr, hop_length=hop_length, tuning=0)
max_response = np.max(np.abs(hcqt), axis=1)
ref_response = np.max(max_response)
continuity = np.abs(np.diff(max_response))
# Test that continuity is never violated by more than 75% point-wise energy
assert np.max(continuity) <= 0.6 * ref_response, np.max(continuity)
# Test that peak-energy deviation is bounded
assert np.std(max_response) < 0.5 * ref_response, np.std(max_response)
def __test(sr, hop_length, y):
hcqt = librosa.hybrid_cqt(y=y, sr=sr, hop_length=hop_length, tuning=0)
max_response = np.max(np.abs(hcqt), axis=1)
ref_response = np.max(max_response)
continuity = np.abs(np.diff(max_response))
# Test that continuity is never violated by more than 75% point-wise energy
assert np.max(continuity) <= 0.6 * ref_response, np.max(continuity)
# Test that peak-energy deviation is bounded
assert np.std(max_response) < 0.5 * ref_response, np.std(max_response)
def _load_music(self, music_file):
y, sr = librosa.load(music_file)
C = librosa.hybrid_cqt(y, sr, fmin=librosa.note_to_hz('C2'), n_bins=72)
CQT = librosa.amplitude_to_db(C, ref=np.max)
mfcc = librosa.feature.mfcc(y=y, sr=sr)[:13, :]
tempo, beats = librosa.beat.beat_track(y=y, sr=sr)
k = 1 + 2 * math.ceil(math.log(len(beats), 2))
C = librosa.util.sync(C, beats) # mean aggregate
mfcc = librosa.util.sync(mfcc, beats)
C = librosa.feature.stack_memory(C)
mfcc = librosa.feature.stack_memory(mfcc)
C_t = C.transpose()
mfcc_t = mfcc.transpose()
return C_t, mfcc_t, k
def analysis(self, save_dir):
wav_name = os.path.splitext(os.path.split(self.wav_dir)[1])[0]
fig = plt.figure(figsize=(50, 10), dpi=100)
fig.tight_layout()
if not os.path.exists(save_dir):
os.mkdir(save_dir)
print('making path ', save_dir)
wav = self.x_wav.reshape(-1)
S = librosa.hybrid_cqt(wav,
fmin=librosa.midi_to_hz(21),
sr=self.sr,
hop_length=128,
bins_per_octave=4 * 12,
n_bins=88 * 4,
filter_scale=0.5)
fig.add_subplot(413)
plt.pcolormesh(self.y_pred_pad + self.notes * 20, cmap='jet')
fig.add_subplot(414)
plt.xlim(-0.5, self.x_wav.shape[0] - 0.5)
plt.plot(range(self.x_wav.shape[0]), self.y_pred_prob, 'ro-')
fig.add_subplot(411)
plt.xlim(-0.5, len(wav) - 0.5)
plt.plot(wav)
fig.add_subplot(412)
plt.pcolormesh(np.abs(S), cmap='jet')
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top=1,
bottom=0,
right=1,
left=0,
hspace=0,
wspace=0)
plt.margins(0, 0)
plt.savefig('{}/{}.jpg'.format(save_dir, wav_name))
plt.show()
plt.clf()
print('saving presion analysis')
def compute_features(self):
"""Actual implementation of the features.
Returns
-------
pcp: np.array(N, F)
The features, each row representing a feature vector for a give
time frame/beat.
"""
audio_harmonic, _ = self.compute_HPSS()
pcp_cqt = np.abs(librosa.hybrid_cqt(audio_harmonic,
sr=self.sr,
hop_length=self.hop_length,
n_bins=self.n_bins,
norm=self.norm,
fmin=self.f_min)) ** 2
pcp = librosa.feature.chroma_cqt(C=pcp_cqt,
sr=self.sr,
hop_length=self.hop_length,
n_octaves=self.n_octaves,
fmin=self.f_min).T
return pcp
def get_features(file, fft=4096, hop=1024, ref=np.max, norm=np.inf):
print(file)
features = {}
y, sr = lb.load(file)
# print('y:', y.shape)
# print('sr:', sr)
tempo, beats = lb.beat.beat_track(y=y, sr=sr, trim=False, hop_length=hop)
beat_track = {'bpm' : tempo, 'beats' : beats.tolist()}
lin_cqt = np.abs(lb.cqt(y=y, sr=sr, hop_length=hop, norm=norm)) ** 2
cqt = lb.amplitude_to_db(lin_cqt, ref=ref)
features['cqt'] = lb.util.sync(cqt, beats).tolist()
lin_cens = np.abs(lb.feature.chroma_cens(y=y, sr=sr, hop_length=hop)) ** 2
cens = lb.amplitude_to_db(lin_cens, ref=ref)
features['cens'] = lb.util.sync(cens, beats).tolist()
harmony, _ = lb.effects.hpss(y=y)
pcp_cqt = np.abs(lb.hybrid_cqt(harmony, sr=sr, hop_length=hop, norm=norm, fmin=27.5)) ** 2
pcp = lb.feature.chroma_cqt(C=pcp_cqt, sr=sr, hop_length=hop, n_octaves=6, fmin=27.5)
features['pcp'] = lb.util.sync(pcp, beats).tolist()
tonnetz = lb.feature.tonnetz(chroma=pcp)
features['tonnetz'] = lb.util.sync(tonnetz, beats).tolist()
mel = lb.feature.melspectrogram(y=y, sr=sr, n_fft=fft, hop_length=hop)
log_mel = lb.amplitude_to_db(mel, ref=ref)
mfcc = lb.feature.mfcc(S=log_mel, n_mfcc=14)
features['mfcc'] = lb.util.sync(mfcc, beats).tolist()
tempogram = lb.feature.tempogram(y=y, sr=sr, hop_length=hop, win_length=192)
features['tempogram'] = lb.util.sync(tempogram, beats).tolist()
return {'beat_track' : beat_track, 'features' : features}
def _calculate_pcp(y, sr):
pcp_cqt = np.abs(librosa.hybrid_cqt(y=y, sr=sr))**2
return librosa.feature.chroma_cqt(C=pcp_cqt, sr=sr)
def analysis(self):
print_scalar = 25
fig_size = 5
fig = plt.figure(figsize=(1.5 * fig_size, 2 * fig_size), dpi=100)
fig.tight_layout()
onset_pred_split = 44
for j, i in enumerate(self.P_false_index):
if i < 26: continue
if (i + 1 + print_scalar) > len(self.y_pred): continue
P_save_dir = 'pic/analysis/{}'.format(
self.input_dir.split('/')[-2])
if not os.path.exists(P_save_dir):
os.mkdir(P_save_dir)
os.mkdir(P_save_dir + '/precision')
hight_light = np.zeros_like(
self.y_onset_pad[:, i - print_scalar:i + 1 + print_scalar])
hight_light[:, print_scalar] = 100
hight_light = hight_light[onset_pred_split:, :]
padding_onset = np.concatenate(
(40 * self.y_onset_pad[:onset_pred_split, i - print_scalar:i +
1 + print_scalar], hight_light +
20 * self.y_pred_pad[onset_pred_split:,
i - print_scalar:i + 1 + print_scalar]),
axis=0)
wav = self.x_wav[i - print_scalar:i + 1 +
print_scalar, :].reshape(-1)
S = librosa.hybrid_cqt(wav,
fmin=librosa.midi_to_hz(21),
sr=self.sr,
hop_length=128,
bins_per_octave=4 * 12,
n_bins=88 * 4,
filter_scale=1)
fig.add_subplot(413)
plt.pcolormesh(
self.y_groundtruth[:, i - print_scalar:i + 1 + print_scalar] +
padding_onset,
vmin=0,
vmax=50,
cmap='jet')
fig.add_subplot(414)
plt.xlim(-0.5, 50.5)
plt.plot(range(51),
self.y_onset[i - print_scalar:i + 1 + print_scalar],
'ro-')
plt.plot(range(51),
self.y_pred_prob[i - print_scalar:i + 1 + print_scalar],
'bo-')
fig.add_subplot(411)
plt.xlim(-0.5, len(wav) - 0.5)
plt.plot(wav)
fig.add_subplot(412)
plt.pcolormesh(np.abs(S), cmap='jet')
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top=1,
bottom=0,
right=1,
left=0,
hspace=0,
wspace=0)
plt.margins(0, 0)
plt.savefig(
'{}/precision/i-{}__time-{:.2f}__prob-{:.2f}__true.jpg'.format(
P_save_dir, i, 440 * i / self.sr, self.y_pred_prob[i]))
# plt.show()
plt.clf()
print('saving presion analysis {}/{}'.format(
j, len(self.P_false_index)),
end='\r')
##########################################################################################################################################################################
##########################################################################################################################################################################
for j, i in enumerate(self.R_false_index):
if i < 26: continue
R_save_dir = 'pic/analysis/{}'.format(
self.input_dir.split('/')[-2])
if not os.path.exists(R_save_dir + '/recall'):
os.mkdir(R_save_dir + '/recall')
hight_light = np.zeros_like(
self.y_onset_pad[:, i - print_scalar:i + 1 + print_scalar])
hight_light[:, print_scalar] = 100
hight_light = hight_light[onset_pred_split:, :]
padding_onset = np.concatenate(
(40 * self.y_onset_pad[:onset_pred_split, i - print_scalar:i +
1 + print_scalar] + hight_light,
20 * self.y_pred_pad[onset_pred_split:,
i - print_scalar:i + 1 + print_scalar]),
axis=0)
wav = self.x_wav[i - print_scalar:i + 1 +
print_scalar, :].reshape(-1)
S = librosa.hybrid_cqt(wav,
fmin=librosa.midi_to_hz(21),
sr=self.sr,
hop_length=128,
bins_per_octave=4 * 12,
n_bins=88 * 4,
filter_scale=1)
fig.add_subplot(413)
plt.pcolormesh(
self.y_groundtruth[:, i - print_scalar:i + 1 + print_scalar] +
padding_onset,
vmin=0,
vmax=50,
cmap='jet')
fig.add_subplot(414)
plt.xlim(-0.5, 50.5)
plt.plot(range(51),
self.y_onset[i - print_scalar:i + 1 + print_scalar],
'ro-')
plt.plot(range(51),
self.y_pred_prob[i - print_scalar:i + 1 + print_scalar],
'bo-')
fig.add_subplot(411)
plt.xlim(-0.5, len(wav) - 0.5)
plt.plot(wav)
fig.add_subplot(412)
plt.pcolormesh(np.abs(S), cmap='jet')
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top=1,
bottom=0,
right=1,
left=0,
hspace=0,
wspace=0)
plt.margins(0, 0)
plt.savefig(
'{}/recall/i-{}__time-{:.2f}__prob-{:.2f}__pred.jpg'.format(
R_save_dir, i, 440 * i / self.sr, self.y_pred_prob[i]))
# plt.show()
plt.clf()
print('saving recall analysis {}/{}'.format(
j, len(self.R_false_index)),
end='\r')
def extract_features(self, y, sr):
try:
if self.params['normalize']:
rms = np.sqrt(np.mean(y * y))
if rms > 1e-4:
y = y / rms
except KeyError:
pass
try:
if self.params['remove_silence']:
y = self.remove_silence(y,
window=32,
hop=32,
threshold=self.params['sil_threshold'])
except KeyError:
pass
if self.params['method'] == 'FFT':
x = np.abs(
librosa.stft(y,
n_fft=self.params['n_fft'],
hop_length=self.params['hop_length']))
x = librosa.logamplitude(x**2)
#x = np.abs(librosa.stft(y,n_fft=320))
elif self.params['method'] == 'MelSpectrogram':
x = librosa.feature.melspectrogram(
y,
sr,
n_fft=self.params['n_fft'],
hop_length=self.params['hop_length'],
n_mels=self.params['n_mels'])
x = librosa.amplitude_to_db(x)
#x = librosa.logamplitude(x**2)
#x = librosa.feature.melspectrogram(y,sr,n_fft=320,n_mels=160)
#
print(x.shape)
raise ValueError
elif self.params['method'] == 'ACF':
numZeros = self.params['hop_length'] - (len(y) %
self.params['hop_length'])
## Use center padding
y2 = np.insert(y, 0, np.zeros(numZeros // 2))
y2 = np.append(y2, np.zeros(numZeros - (numZeros // 2)))
ind = 0
blockSize = self.params['n_fft']
while ((ind + blockSize) <= len(y2)):
if ind == 0:
x = librosa.autocorrelate(
y2[ind:ind + blockSize],
max_size=(self.params['hop_length'] // 2))
x = np.expand_dims(x, 0)
else:
x = np.vstack(
(x,
librosa.autocorrelate(
y2[ind:ind + blockSize],
max_size=(self.params['hop_length'] // 2))))
ind += blockSize
#x = x.transpose()
x = resample(x, 96, t=None, axis=0)
#print(x.shape)
elif self.params['method'] == 'Cepstrum':
x = librosa.stft(y,
n_fft=self.params['n_fft'],
hop_length=self.params['hop_length'])
x = np.log(x)
for i in range(len(x)):
x[i] = np.absolute(np.fft.ifft(x[i]))
x = x.real.astype('float32')
x = resample(x, 96, t=None, axis=0)
elif self.params['method'] == 'CQT':
#x = librosa.hybrid_cqt(y,sr, hop_length=self.params['hop_length'], n_bins=self.params['n_bins'], bins_per_octave=self.params['bins_per_octave'])
x = librosa.hybrid_cqt(y,
sr,
hop_length=128,
n_bins=144,
bins_per_octave=24)
x = librosa.amplitude_to_db(x**2)
elif self.params['method'] == 'MFCC':
x = librosa.feature.mfcc(y,
sr,
n_fft=self.params['n_fft'],
n_mels=self.params['n_mels'],
n_mfcc=self.params['n_mfcc'],
hop_length=self.params['hop_length'])
delta = librosa.feature.delta(x)
d_delta = librosa.feature.delta(x, order=2)
x = np.concatenate([x, delta, d_delta], axis=0)
return torch.FloatTensor(x)
cqt = librosa.cqt(y, sr)
cqt = np.abs(cqt)
cqt = cqt.astype(np.float32)
print(cqt.shape, cqt.dtype)
d_cqt = librosa.amplitude_to_db(cqt, ref=np.max)
librosa.display.specshow(d_cqt,
y_axis='log',
x_axis='time',
sr=sr,
cmap='viridis')
plt.colorbar(format='%+2.0f dB')
plt.title('cqt-spectrogram')
plt.show()
### hybrid-cqt-spectrogram (混合常量Q变换) (*84, t) *n_bins
hcqt = librosa.hybrid_cqt(y=y, sr=sr)
hcqt = np.abs(hcqt)
hcqt = hcqt.astype(np.float32)
print(hcqt.shape, hcqt.dtype)
d_hcqt = librosa.amplitude_to_db(hcqt, ref=np.max)
librosa.display.specshow(d_hcqt,
y_axis='log',
x_axis='time',
sr=sr,
cmap='viridis')
plt.colorbar(format='%+2.0f dB')
plt.title('hybrid-cqt-spectrogram')
plt.show()
### pseudo-cqt-spectrogram (伪常量Q变换) (*84, t) *n_bins
pcqt = librosa.pseudo_cqt(y=y, sr=sr)