def test_vqt(
y_cqt,
sr_cqt,
hop_length,
fmin,
n_bins,
gamma,
bins_per_octave,
tuning,
filter_scale,
norm,
res_type,
sparsity,
):
C = librosa.vqt(
y=y_cqt,
sr=sr_cqt,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
gamma=gamma,
bins_per_octave=bins_per_octave,
tuning=tuning,
filter_scale=filter_scale,
norm=norm,
sparsity=sparsity,
res_type=res_type,
)
# type is complex
assert np.iscomplexobj(C)
# number of bins is correct
assert C.shape[0] == n_bins
def AMT_Framing(filename_):
# Audio Processing
#filename = 'Guns N Roses-Sweet Child O Mine Intro.wav'
filename = "{}".format(filename_)
x, fs = librosa.load(filename, sr=None, mono=True, duration=DURATION)
V = librosa.vqt(x,
sr=fs,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
gamma=20,
bins_per_octave=bins_per_octave,
tuning=tuning,
filter_scale=filter_scale,
norm=norm,
sparsity=0.01,
window='hann',
scale=scale,
pad_mode=pad_mode,
res_type=res_type,
dtype=dtype)
# Mapping Magnitude spectrogram to the Mel Scale
V_mel = np.abs(V)
logFrame = librosa.amplitude_to_db(V_mel)
mels = librosa.feature.melspectrogram(S=V_mel,
sr=fs,
n_mels=n_mels,
n_fft=hop_length * 2,
hop_length=hop_length)
np_array_list = []
np_array_list.append(mels)
frame_windows_list = []
numSlices_list = []
Y_numSlices = 625
for i in range(len(np_array_list)):
VQT_result = np_array_list[i]
paddedX = np.zeros(
(VQT_result.shape[0], VQT_result.shape[1] + WINDOW_SIZE - 1),
dtype=float)
pad_amount = int(WINDOW_SIZE / 2)
paddedX[:, pad_amount:-pad_amount] = VQT_result
frame_windows = np.array([
paddedX[:, j:j + WINDOW_SIZE] for j in range(VQT_result.shape[1])
])
#frame_windows = np.expand_dims(frame_windows, axis=3)
numSlices = min(frame_windows.shape[0], Y_numSlices)
numSlices_list.append(numSlices)
frame_windows_list.append(frame_windows[:numSlices])
audio_frames = np.concatenate(frame_windows_list, axis=0)
#audio_frames = frame_windows_list
#storingData(audio_frames)
#print("Windows shape: ",audio_frames.shape)
return audio_frames
def AMT(filename_):
# Audio Processing
# Loading the Audios
# Path Configuration
#path = os.getcwd() + '/' + filename_
filename = "{}".format(filename_)
x, fs = librosa.load(filename, sr=None, mono=True, duration=12)
# VQT Computation
V = librosa.vqt(x, sr=fs, hop_length=hop_length, fmin=fmin, n_bins=n_bins, gamma=20, bins_per_octave=bins_per_octave, tuning=tuning,
filter_scale=filter_scale, norm=norm, sparsity=0.01, window='hann', scale=scale, pad_mode=pad_mode, res_type=res_type, dtype=dtype)
V_mel = np.abs(V)
# Conversion into the Mel-Scale to display and save Mel-spectrogram for prediction
melspec(V_mel,filename)
def test_vqt(
y_cqt_110,
sr_cqt,
hop_length,
fmin,
n_bins,
gamma,
bins_per_octave,
tuning,
filter_scale,
norm,
res_type,
sparsity,
):
C = librosa.vqt(
y=y_cqt_110,
sr=sr_cqt,
hop_length=hop_length,
fmin=fmin,
n_bins=n_bins,
gamma=gamma,
bins_per_octave=bins_per_octave,
tuning=tuning,
filter_scale=filter_scale,
norm=norm,
sparsity=sparsity,
res_type=res_type,
)
# type is complex
assert np.iscomplexobj(C)
# number of bins is correct
assert C.shape[0] == n_bins
if fmin is None:
fmin = librosa.note_to_hz("C1")
# check for peaks if 110 is within range
if 110 <= fmin * 2 ** (n_bins / bins_per_octave):
peaks = np.argmax(np.abs(C), axis=0)
# This is our most common peak index in the CQT spectrum
# we use the mode here over frames to sidestep transient effects
# at the beginning and end of the CQT
common_peak = scipy.stats.mode(peaks)[0][0]
# Convert peak index to frequency
peak_frequency = fmin * 2 ** (common_peak / bins_per_octave)
def VQT_from_file(audio_file, bins_per_octave=60, n_octaves=8, gamma=20):
y, fs = librosa.load(audio_file, sr=25600)
vqt = librosa.vqt(y,
sr=fs,
hop_length=256,
fmin=SpectrogramUtil.FMIN,
n_bins=bins_per_octave * n_octaves,
bins_per_octave=bins_per_octave,
gamma=gamma)
log_vqt = (
(1. / 80.) *
librosa.amplitude_to_db(np.abs(np.array(vqt)), ref=np.max)) + 1.
return log_vqt
def process_mel_spectrogram(self, n_mels, CQT=False, VQT=False):
if CQT:
C = np.abs(librosa.cqt(self.audio_waveform, sr=self.sample_rate, n_bins=112))
CQT = librosa.amplitude_to_db(C, ref=np.max)
self.mel_spectrogram = np.flipud(CQT)
elif VQT:
V = np.abs(librosa.vqt(self.audio_waveform, sr=self.sample_rate, n_bins=112))
VQT = librosa.amplitude_to_db(V, ref=np.max)
self.mel_spectrogram = np.flipud(VQT)
else:
mel_spectrogram = np.flipud(librosa.feature.melspectrogram(self.audio_waveform, sr=self.sample_rate, n_mels=n_mels))
log_mel_spectrogram = librosa.power_to_db(mel_spectrogram, ref=np.max)
self.mel_spectrogram = log_mel_spectrogram
midi_time = self.midi_note_array.shape[1]*self.tempo/self.PPQ/1000000
audio_time = self.mel_spectrogram.shape[1]*512/self.sample_rate
# Fixing audio and midi discrepancy
time_difference = audio_time - midi_time
time_difference_sample_ticks = round(time_difference*self.sample_rate/512)
print(time_difference_sample_ticks)
if time_difference_sample_ticks > 0:
self.mel_spectrogram = self.mel_spectrogram[:, 0:-time_difference_sample_ticks]
self.song_total_sample_ticks = self.mel_spectrogram.shape[1]
ax.set_ylabel('log Hz', size=15)
ax.set_xlabel('Time', size=15)
plt.show()
plt.show()
mel_spectrogram = librosa.feature.melspectrogram(y, sr=sr, n_mels=128)
print(mel_spectrogram)
log_mel_sprectrogram = librosa.power_to_db(mel_spectrogram)
C = np.abs(librosa.cqt(y, sr=sr, n_bins=112))
CQT = librosa.amplitude_to_db(C, ref=np.max)
print(CQT)
V = np.abs(librosa.vqt(y, sr=sr, n_bins=112))
VQT = librosa.amplitude_to_db(V, ref=np.max)
print(CQT)
X = librosa.stft(y)
Xdb = librosa.amplitude_to_db(abs(X))
fig, axs = plt.subplots(1, 3, figsize=(15, 20))
axs[0].imshow(np.flipud(VQT), aspect='auto', interpolation='nearest')
axs[1].imshow(np.flipud(log_mel_sprectrogram),
aspect='auto',
interpolation='nearest')
axs[2].imshow(np.flipud(CQT), aspect='auto', interpolation='nearest')
plt.show()
# fig, axs = plt.subplots()
# plt.grid(b=None)
def librosa_hvqt(audio, harmonics, sample_rate, hop_length, fmin, n_bins,
bins_per_octave, gamma):
"""
Compute an HVQT using Librosa.
Parameters
----------
audio : ndarray (N)
Audio to transform,
N - number of samples
harmonics : list of ints
Specific harmonics to stack across the harmonic dimension
sample_rate : int or float
Number of samples per second of audio
hop_length : int
Number of samples between frames
fmin : float
Lowest center frequency in basis
n_bins : int
Number of basis functions in the filterbank
bins_per_octave : int
Number of basis functions per octave
gamma : float
Bandwidth offset to smoothly vary Q-factor
Returns
----------
hvqt : ndarray (H x F x T)
Harmonic Variable-Q Transform (HVQT) for the provided audio,
H - number of harmonics
F - number of bins
T - number of time steps (frames)
"""
# Initialize a list to hold the harmonic-wise transforms
hvqt = list()
# Initialize a list to hold the number of frames for each transform
frames = list()
# Loop through harmonics
for h in range(len(harmonics)):
# Compute the true minimum center frequency for this harmonic
h_fmin = harmonics[h] * fmin
# Compute the VQT for this harmonic
vqt = librosa.vqt(audio,
sr=sample_rate,
hop_length=hop_length,
fmin=h_fmin,
n_bins=n_bins,
gamma=gamma,
bins_per_octave=bins_per_octave)
# Keep track of the number of frames produced
frames.append(vqt.shape[-1])
# Add the VQT to the collection
hvqt.append(np.expand_dims(vqt, axis=0))
# Determine the maximum number of frames that can be concatenated
max_frames = min(frames)
# Perform any trimming and concatenate
hvqt = np.concatenate([vqt[..., :max_frames] for vqt in hvqt])
# Take the magnitude
hvqt = np.abs(hvqt)
return hvqt
