def expe_2():
sig_1_path = '/sons/voxforge/main/Learn/cmu_us_slt_arctic/wav/arctic_a0372.wav'
sig_2_path = '/sons/voxforge/main/Learn/cmu_us_rms_arctic/wav/arctic_a0372.wav'
i = 0
for sig_path in [sig_1_path, sig_2_path]:
synth_sig = Signal(sig_path, normalize=True, mono=True)
#synth_sig.crop(0.1*synth_sig.fs, 3.5*synth_sig.fs)
#synth_sig.resample(32000)
plt.figure(figsize=(10, 10))
plt.subplot(211)
plt.plot(
np.arange(.0, synth_sig.length) / float(synth_sig.fs),
synth_sig.data)
plt.xticks([])
plt.ylim([-1, 1])
plt.grid()
plt.subplot(212)
synth_sig.spectrogram(1024,
64,
order=0.5,
log=False,
cmap=cm.hot,
cbar=False)
plt.savefig(op.join(figure_output_path, 'voice_%d_spectro.pdf' % i))
synth_sig.write(op.join(audio_output_path, 'voice_%d_spectro.wav' % i))
i += 1
def recons_save_fig_audio(magspec,
target_name,
n_max_frames,
fs=22050,
format=(8, 3),
nb_gl_iter=30):
init_vec = np.random.randn(128 * n_max_frames)
x_recon = transforms.gl_recons(magspec[:, :n_max_frames],
init_vec,
nb_gl_iter,
512,
128,
display=False)
rec_sig = Signal(x_recon, fs, normalize=True)
rec_sig.write(os.path.join(output_audio_path, '%s.wav' % target_name))
plt.figure(figsize=format)
rec_sig.spectrogram(512, 128, order=1, log=True, cmap=cm.jet, cbar=False)
plt.savefig(os.path.join(output_fig_path, '%s.png' % target_name))
def expe_1():
synth_sig = Signal(audio_test_file, normalize=True, mono=True)
synth_sig.crop(0.1 * synth_sig.fs, 3.5 * synth_sig.fs)
#synth_sig.resample(32000)
plt.figure(figsize=(10, 5))
plt.subplot(211)
plt.plot(
np.arange(.0, synth_sig.length) / float(synth_sig.fs), synth_sig.data)
plt.xticks([])
plt.ylim([-1, 1])
plt.grid()
plt.subplot(212)
synth_sig.spectrogram(1024,
64,
order=0.25,
log=False,
cmap=cm.hot,
cbar=False)
plt.savefig(op.join(figure_output_path, 'glocs_spectro.pdf'))
plt.show()
sk = STFTPeaksSketch(**{'scale': 2048, 'step': 512})
orig_sig = Signal(audio_test_file, normalize=True, mono=True)
noisy_sig = Signal(orig_sig.data + 0.2 * np.random.randn(orig_sig.length),
orig_sig.fs,
normalize=True,
mono=True)
sk.recompute(orig_sig)
sk.sparsify(20)
plt.figure(figsize=(10, 6))
plt.subplot(221)
orig_sig.spectrogram(512,
128,
order=2,
log=True,
ax=plt.gca(),
cmap=cm.bone_r,
cbar=False)
plt.subplot(222)
noisy_sig.spectrogram(512,
128,
order=2,
log=True,
ax=plt.gca(),
cmap=cm.bone_r,
cbar=False)
plt.subplot(223)
fgpthandle._build_pairs(sk.fgpt(), sk.params, display=True, ax=plt.gca())
plt.gca().invert_yaxis()
plt.yticks([])
output_fig_path = '/home/manu/Documents/Articles/ISMIR2013/ListeningMSD/Figures/'
colormap = cm.jet
format = (8,3)
# also load the Dan Ellis's synthesized version
# The Piano cross-synthesis and the Viterbi smoothed Musaicing?
# resynthesize using the first N frames
n_max_frames = 900
nb_gl_iter = 30
init_vec = np.random.randn(128*n_max_frames)
x_recon_median = transforms.gl_recons(median_magspec[:,:n_max_frames], init_vec, nb_gl_iter,
512, 128, display=False)
sig_median = Signal(x_recon_median, 22050,normalize=True)
sig_median.write(os.path.join(output_audio_path, '%s_add_median.wav'%t_name))
plt.figure(figsize=format)
sig_median.spectrogram(512, 128, order=1, log=True, cmap=colormap, cbar=False)
plt.savefig(os.path.join(output_fig_path, '%s_add_median.png'%t_name))
init_vec = np.random.randn(128*n_max_frames)
x_recon_orig = transforms.gl_recons(orig_spec[:,:n_max_frames], init_vec, nb_gl_iter,
512, 128, display=False)
sig_orig= Signal(x_recon_orig, 22050,normalize=True)
sig_orig.write(os.path.join(output_audio_path, '%s_original.wav'%t_name))
plt.figure(figsize=format)
sig_orig.spectrogram(512, 128, order=1, log=True, cmap=colormap, cbar=False)
plt.savefig(os.path.join(output_fig_path, '%s_original.png'%t_name))
init_vec = np.random.randn(128*n_max_frames)
x_recon_max = transforms.gl_recons(max_magspec[:,:n_max_frames], init_vec, nb_gl_iter,
512, 128, display=False)
sig_max= Signal(x_recon_max, 22050,normalize=True)
'''
doc.pyplots.Spectro_example - Created on Apr 23, 2013
@author: M. Moussallam
'''
import os.path as op
from PyMP import Signal, approx
import matplotlib.pyplot as plt
abPath = op.abspath('../../data/')
sig = Signal(op.join(abPath, 'glocs.wav'), normalize=True, mono=True)
import matplotlib.cm as cm
plt.figure()
sig.spectrogram(1024, 128, order=2, log=True, cmap=cm.hot, cbar=True)
plt.show()
print t_path + '/' + t_name + '.au'
from feat_invert.transforms import get_audio
orig, fs = get_audio(t_path + '/' + t_name + '.au',
0,
np.sum(t_seg_duration[:nb_max_seg]) + marge,
targetfs=22050)
#orig = Signal(t_path + t_name + '.au', normalize=True)
print "Working on %s duration of %2.2f" % (
t_name, np.sum(t_seg_duration[:nb_max_seg]))
orig_spec = np.abs(stft.stft(orig, 512, 128)[0, :, :])
sig_ellis = Signal('%sellis_resynth%s.wav' % (output_audio_path, t_name),
normalize=True)
#sig_ellis.crop(0,orig_spec.shape[1]*128 + 512-128)
plt.figure(figsize=(8, 3))
sig_ellis.spectrogram(512, 128, order=1, log=True, cmap=cm.jet, cbar=False)
plt.savefig(os.path.join(output_fig_path, '%s_ellis.png' % t_name))
feat_idxs = get_feat_idx_from_name(feat_combs[0])
# limit the size of the l_features vector:
l_feats = l_feats_all[:M, feat_idxs]
# and the t_feats vector
t_feats = t_feats_all[:nb_max_seg, feat_idxs]
# standardizing the features
for f in range(l_feats.shape[1]):
mu = np.mean(l_feats[:, f])
sigma = np.std(l_feats[:, f])
# print "Mu %2.4f Sigma: %2.4f"%(mu,sigma)
l_feats[:, f] = (l_feats[:, f] - mu) / sigma
t_feats[:, f] = (t_feats[:, f] - mu) / sigma
import matplotlib.colors as cc
import matplotlib.cm as cm
from PyMP import Signal
import stft
from scipy.signal import lfilter, hann
plt.switch_backend('Agg')
audio_test_file = '/home/manu/workspace/recup_angelique/Sketches/NLS Toolbox/Hand-made Toolbox/forAngelique/61_sadness.wav'
audio_test_file = '/sons/jingles/panzani.wav'
figure_output_path = '/home/manu/workspace/audio-sketch/src/reporting/figures/'
sig = Signal(audio_test_file, mono=True, normalize=True)
sig.downsample(16000)
scale = 512
step = 32
sig.spectrogram(scale, step, order=1, log=False, cmap=cm.coolwarm)
def plot_spectrogram(sig_stft, scale=512, step=128):
plt.figure()
plt.imshow(20 * np.log10(np.abs(sig_stft[0, :, :])),
aspect='auto',
origin='lower',
interpolation='nearest',
cmap=cm.copper_r)
x_tick_vec = (np.linspace(0, sig_stft.shape[2], 10)).astype(int)
x_label_vec = (x_tick_vec * float(step)) / float(sig.fs)
y_tick_vec = (np.linspace(0, sig_stft.shape[1], 6)).astype(int)
y_label_vec = (y_tick_vec / float(scale)) * float(sig.fs)
name="unnamed")
plt.figure()
plt.plot(x_recon_viterbi - x_recon_1nn)
plt.show()
# is the reconstructed spectrogram relevantly statistically distributed ?
import matplotlib.cm as cm
from scipy.ndimage import median_filter
plt.figure()
plt.subplot(311)
plt.imshow(np.log10(median_filter(Y_hat.T, [3, 1])), origin='lower')
plt.subplot(312)
plt.imshow(np.log10(Y_hat_viterbi.T), origin='lower')
plt.plot([i * 2 for i in best_path], 'k')
plt.subplot(313)
orig_sig.spectrogram(win_size, 128, order=1, log=True, cmap=cm.jet, cbar=False)
plt.show()
filt_sizes = [[1, 1], [3, 1], [5, 1], [10, 1], [1, 3], [1, 5], [1, 10], [3, 3],
[5, 5], [10, 10]]
x_recons = []
for filt_size in filt_sizes:
guess_yhat = median_filter(Y_hat_viterbi, filt_size)
init_vec = np.random.randn(128 * guess_yhat.shape[0])
x_recons.append(
transforms.gl_recons(guess_yhat.T,
init_vec,
50,
win_size,
128,
display=False))
