def save(self,
save_audio=False,
save_array=True,
save_image=True,
save_info=True,
naming_by="midi_number"):
# The output_name for the saving data
if naming_by == "midi_number":
output_name = str(self.note_number) + "_" + str(round(self.duration, 3)) \
+ "_" + str(self.velocity)
elif naming_by == "nameWithOctave":
output_name = self.pitch.nameWithOctave + "_" + str(round(self.duration, 3)) \
+ "_" + str(self.velocity)
else:
raise Exception("Parameter naming_by not understood.")
# Save the audio
if save_audio:
wav.write(
Path(SAMPLES_AUDIO_PATH) / Path(output_name + '.wav'), FS,
self.signal)
# Save array
if save_array:
np.save(Path(SAMPLES_ARRAYS_PATH) /
Path(output_name + '_spectrogram' + '.npy'),
self.spectrogram_log,
allow_pickle=True)
# Save image
if save_image:
plot_cqt(self.spectrogram_log,
self.time_vector,
fig_title="Sample " + str(self),
show=False)
plt.savefig(Path(SAMPLES_IMAGES_PATH) / Path(output_name + '.png'),
dpi=DPI,
format='png')
plt.close()
# Save info
if save_info:
np.save(Path(SAMPLES_INFO_PATH) /
Path(output_name + '_distribution' + '.npy'),
self.partials_distribution.linear_regressions,
allow_pickle=True)
np.save(Path(SAMPLES_INFO_PATH) /
Path(output_name + '_bins' + '.npy'),
self.partials_bins,
allow_pickle=True)
np.save(Path(SAMPLES_INFO_PATH) /
Path(output_name + '_amplitudes' + '.npy'),
self.partials_amplitudes,
allow_pickle=True)
def example_1():
_sta = time.time()
_signal = signal_from_file('anastasia')
_end = time.time()
log.info("Time to recover signal: " + str(round(_end - _sta, 3)) + " seconds.")
_sta = time.time()
_spectrogram, _time_vector = cqt(_signal)
_end = time.time()
log.info("Time compute the CQT: " + str(round(_end - _sta, 3)) + " seconds.")
_sta = time.time()
plot_cqt(_spectrogram, _time_vector)
_end = time.time()
log.info("Time to plot: " + str(round(_end - _sta, 3)) + " seconds.")
harmonic=True,
frequency_decay_dependency=decay)
samples_set = SamplesSet.from_synthesis(partials_distribution)
return samples_set
else:
raise ValueError('Invalid parameter samples_type.')
if __name__ == '__main__':
_samples_set = get_samples_set('basic')
_piece = midi2piece('tempest_3rd-start')
_signal = _samples_set.synthesize(_piece)
_spectrogram, _time_vector = cqt(_signal)
plot_cqt(_spectrogram, _time_vector)
play = False
if play:
sd.play(_signal, FS)
# _samples_name = 'samples'
# _instrument = "MyPiano"
#
# # _samples_set = SamplesSet.from_directory("MyPiano", "samples", load_all=LOAD_ALL)
# _samples_set = SamplesSet.from_midi_file("MyPiano", "samples", save_audio=True, save_array=True,
# save_image=True, save_info=True)
def cqt(signal, numpy=True, db=True):
time_array = np.arange(np.ceil(
signal.size / HOP_LENGTH).astype(int)) / (FS / HOP_LENGTH)
signal_tensor = torch.tensor(signal, device=DEVICE, dtype=torch.float)
cqt_tensor = cqt_layer(signal_tensor, normalization_type='wrap')
if db:
cqt_tensor = 20 * torch.log10(cqt_tensor + EPS)
if numpy:
cqt_array = cqt_tensor.cpu().numpy()[0, :, :]
torch.cuda.empty_cache()
return cqt_array, time_array
else:
# time_tensor = torch.tensor(time_array)
return cqt_tensor[0, :, :], time_array
if __name__ == '__main__':
_signal = np.zeros(FS * 2)
dirac_width = 1
for i in range(dirac_width):
_signal[i::int(FS * 0.033)] = 1
_signal += 0.1 * np.sin(2 * np.pi * 440 * np.arange(FS * 2) / FS)
_spectrogram, _time_array = cqt(_signal)
plot_cqt(_spectrogram, _time_array, v_min=-100)
sample_name = "A2_12.787_113"
# Partials
fundamental_bin = 2 * BINS_PER_OCTAVE
partials_pos = fundamental_bin + np.round(
np.log2(np.arange(N_PARTIALS) + 1) * BINS_PER_OCTAVE).astype(int)
start = 2 # in seconds
end = 6 # in seconds
_signal = signal_from_file(sample_name, SAMPLES_AUDIO_PATH)
_spectrogram, _time_vector = cqt(
_signal, numpy=True)[:,
np.floor(start / TIME_RESOLUTION).astype(int):np.
ceil(end / TIME_RESOLUTION).astype(int)]
plot_cqt(_spectrogram, _time_vector, fig_title=sample_name)
a_erosion_1 = erode(_spectrogram, _strel_1, _origin_1)
plot_cqt(a_erosion_1,
_time_vector,
fig_title="Erosion - 1",
v_min=-40,
c_map='Greys')
# Spectrogram
partials_amplitudes = _spectrogram[partials_pos, :]
partials_distribution = partials_amplitudes[:, 0] - np.max(
partials_amplitudes[:, 0])
plt.figure()
for i in range(len(partials_pos)):
if __name__ == '__main__':
# Parameters
play = True
# Create the signal
sta = time.time()
samples_set = get_samples_set('basic')
end = time.time()
log.info("Time to create samples set: " + str(round(end - sta, 3)) + " seconds.")
sta = time.time()
piece = midi2piece('prelude_em')
signal = samples_set.synthesize(piece)
end = time.time()
log.info("Time to synthesize the signal: " + str(round(end - sta, 3)) + " seconds.")
# Time-frequency transform of the signal
sta = time.time()
spectrogram, time_vector = cqt(signal)
end = time.time()
log.info("Time to compute the CQT of the signal: " + str(round(end - sta, 3)) + " seconds.")
if play:
sd.play(signal, FS)
plot_cqt(spectrogram, time_vector)
# Morphological transform of the signal
from time_frequency import cqt
import sounddevice as sd
from parameters import *
from utils import to_db, db_to_velocity
from librosa import hz_to_midi
from tqdm import tqdm
from nnMorpho.operations import erosion, dilation, opening, closing
samples_set = get_samples_set('basic')
piece = midi2piece('prelude_em')
signal = samples_set.synthesize(piece)
spectrogram, time_vector = cqt(signal, numpy=False)
plot_cqt(spectrogram, time_vector, numpy=False)
# sd.play(signal, FS)
# Leakage
structural_element_leakage = torch.tensor([[-5.8], [0], [-6]], device=DEVICE)
origin_leakage = (1, 0)
erosion_leakage = erosion(spectrogram,
structural_element_leakage,
origin_leakage,
border_value='euclidean')
plot_cqt(erosion_leakage,
time_vector,
fig_title="Erosion of the leakage",
numpy=False)