return arr
real_mel = reconstruct(real_tf.cluster_centers_, X[0][:, :n_clusters],
n_features)
imag_mel = reconstruct(imag_tf.cluster_centers_, X[0][:, -n_clusters:],
n_features)
rPxx = invmelspec(real_mel, fs, n_fft)
iPxx = invmelspec(imag_mel, fs, n_fft)
Pxx = rPxx + 1j * iPxx
Xs = istft(Pxx).astype('float32')
lb = np.percentile(Xs, 0.5)
ub = np.percentile(Xs, 99.5)
Xs[Xs < lb] = 0.
Xs[Xs > ub] = 0.
Xs -= Xs.mean()
wavfile.write('orig.wav', fs, soundsc(Xs))
model = test_rnnrbm(X, num_epochs=200)
n_samples_to_gen = 10
for i in range(n_samples_to_gen):
g = model.generate()
real_mel = reconstruct(real_tf.cluster_centers_, g[:, :n_clusters],
n_features)
imag_mel = reconstruct(imag_tf.cluster_centers_, g[:, -n_clusters:],
n_features)
rPxx = invmelspec(real_mel, fs)
iPxx = invmelspec(imag_mel, fs)
Pxx = rPxx + 1j * iPxx
Xs = istft(Pxx).astype('float32')
lb = np.percentile(Xs, 0.5)
ub = np.percentile(Xs, 99.5)
if prev_code is not None:
# If there are none active just use the previous
code = prev_code
else:
# Very first sample is messed up... just pick one
code = codebook[0]
arr[i] = code
prev_code = code
return arr
for i in [0, 20, 40, 60, 80, 100]:
lsf = reconstruct(lsf_tf.cluster_centers_, X[i][:, :n_lsf_clusters],
n_features)
a = lsf_to_lpc(lsf)
gain = reconstruct(gain_tf.cluster_centers_, X[i][:, -n_gain_clusters:],
1)
X_s = lpc_synthesis(a, gain, window_step=window_step)
wavfile.write('orig_%i.wav' % i, fs, soundsc(X_s))
model = test_lstmrbm(X, num_epochs=200)
n_samples_to_gen = 5
for i in range(n_samples_to_gen):
g = model.generate()
lsf = reconstruct(lsf_tf.cluster_centers_, g[:, :n_lsf_clusters],
n_features)
a = lsf_to_lpc(lsf)
gain = reconstruct(gain_tf.cluster_centers_, g[:, -n_gain_clusters:],
1)
X_s = lpc_synthesis(a, gain, window_step=window_step)
wavfile.write('sample_%i.wav' % i, fs, soundsc(X_s))
code = prev_code
else:
code = codebook[0]
arr[i] = code
prev_code = code
return arr
print len(X)
for i in [0, 3, 5, 8, 13]:
lsf = reconstruct(lsf_tf.cluster_centers_, X[i][:, :n_lsf_clusters],
n_features)
a = lsf_to_lpc(lsf)
gain = reconstruct(gain_tf.cluster_centers_,
X[i][:, -n_gain_clusters:], 1)
X_s = lpc_synthesis(a, gain, window_step=window_step)
wavfile.write('orig_%i.wav' % i, fs, soundsc(X_s))
model = test_rnnrbm(X, num_epochs=200)
n_samples_to_gen = 5
for i in range(n_samples_to_gen):
g = model.generate()
#f = open('sample_peach_pickle_rnndbn_%i.pkl' % i, 'w')
#cPickle.dump(g,f)
#f.close()
lsf = reconstruct(lsf_tf.cluster_centers_, g[:, :n_lsf_clusters],
n_features)
a = lsf_to_lpc(lsf)
gain = reconstruct(gain_tf.cluster_centers_, g[:, -n_gain_clusters:],
1)
X_s = lpc_synthesis(a, gain, window_step=window_step)
wavfile.write('sample_apple_pineapple_rnndbn_%i.wav' % i, fs,
if prev_code is not None:
code = prev_code
else:
code = codebook[0]
arr[i] = code
prev_code = code
return arr
print len(X)
for i in [0, 3, 5, 8, 13]:
lsf = reconstruct(lsf_tf.cluster_centers_, X[i][:, :n_lsf_clusters],
n_features)
a = lsf_to_lpc(lsf)
gain = reconstruct(gain_tf.cluster_centers_, X[i][:, -n_gain_clusters:],
1)
X_s = lpc_synthesis(a, gain, window_step=window_step)
wavfile.write('orig_%i.wav' % i, fs, soundsc(X_s))
model = test_rnnrbm(X, num_epochs=200)
n_samples_to_gen = 5
for i in range(n_samples_to_gen):
g = model.generate()
#f = open('sample_peach_pickle_rnndbn_%i.pkl' % i, 'w')
#cPickle.dump(g,f)
#f.close()
lsf = reconstruct(lsf_tf.cluster_centers_, g[:, :n_lsf_clusters],
n_features)
a = lsf_to_lpc(lsf)
gain = reconstruct(gain_tf.cluster_centers_, g[:, -n_gain_clusters:],
1)
X_s = lpc_synthesis(a, gain, window_step=window_step)
wavfile.write('sample_apple_pineapple_rnndbn_%i.wav' % i, fs, soundsc(X_s))
