def load_wav_as_mfcc_delta(fname,
target_size=None,
noise_files=None,
augment_with_noise=False,
class_dir=None):
(fs, signal) = utils.read_wave_file(fname)
if class_dir:
signal = da.same_class_augmentation(signal, class_dir)
if augment_with_noise:
signal = da.noise_augmentation(signal, noise_files)
mfcc = librosa.feature.mfcc(signal, fs, n_mfcc=target_size[0])
mfcc_delta_3 = librosa.feature.delta(mfcc, width=3, order=1)
mfcc_delta_11 = librosa.feature.delta(mfcc, width=11, order=1)
mfcc_delta_19 = librosa.feature.delta(mfcc, width=19, order=1)
if target_size:
mfcc = scipy.misc.imresize(mfcc, target_size)
mfcc_delta_3 = scipy.misc.imresize(mfcc_delta_3, target_size)
mfcc_delta_11 = scipy.misc.imresize(mfcc_delta_11, target_size)
mfcc_delta_19 = scipy.misc.imresize(mfcc_delta_19, target_size)
mfcc = mfcc.reshape(mfcc.shape[0], mfcc.shape[1], 1)
mfcc_delta_3 = mfcc_delta_3.reshape(mfcc_delta_3.shape[0],
mfcc_delta_3.shape[1], 1)
mfcc_delta_11 = mfcc_delta_11.reshape(mfcc_delta_11.shape[0],
mfcc_delta_11.shape[1], 1)
mfcc_delta_19 = mfcc_delta_19.reshape(mfcc_delta_19.shape[0],
mfcc_delta_19.shape[1], 1)
mfcc_delta = np.concatenate(
[mfcc, mfcc_delta_3, mfcc_delta_11, mfcc_delta_19], axis=2)
return mfcc_delta
def noise_augmentation_from_dirs(noise_dir, class_dir):
sig_paths = glob.glob(os.path.join(class_dir, "*.wav"))
sig_path = np.random.choice(sig_paths, 1, replace=False)[0]
(fs, sig) = utils.read_wave_file(sig_path)
aug_sig = da.noise_augmentation(sig, noise_dir)
spectrogram_sig = sp.wave_to_sample_spectrogram(sig, fs)
spectrogram_aug_sig = sp.wave_to_sample_spectrogram(aug_sig, fs)
fig = plt.figure(1)
cmap = plt.cm.get_cmap('jet')
gs = gridspec.GridSpec(2, 1)
# whole spectrogram
ax1 = fig.add_subplot(gs[0, 0])
ax1.pcolormesh(spectrogram_sig, cmap=cmap)
ax1.set_title("Original Signal")
ax2 = fig.add_subplot(gs[1, 0])
ax2.pcolormesh(spectrogram_aug_sig, cmap=cmap)
ax2.set_title("Noise Augmented signal")
gs.update(wspace=0.5, hspace=0.5)
basename = utils.get_basename_without_ext(sig_path)
fig.savefig(basename + "_noise_augmentation.png")
fig.clf()
plt.close(fig)
def load_wav_as_tempogram(fname,
target_size=None,
noise_files=None,
augment_with_noise=False,
class_dir=None):
(fs, signal) = utils.read_wave_file(fname)
if class_dir:
signal = da.same_class_augmentation(signal, class_dir)
if augment_with_noise:
signal = da.noise_augmentation(signal, noise_files)
tempogram = sp.wave_to_tempogram(signal, fs)
if target_size:
tempogram = scipy.misc.imresize(tempogram, target_size)
tempogram = tempogram.reshape((tempogram.shape[0], tempogram.shape[1], 1))
return tempogram
def load_wav_as_mfcc(fname,
target_size=None,
noise_files=None,
augment_with_noise=False,
class_dir=None):
(fs, signal) = utils.read_wave_file(fname)
if class_dir:
signal = da.same_class_augmentation(signal, class_dir)
if augment_with_noise:
signal = da.noise_augmentation(signal, noise_files)
mfcc = librosa.feature.mfcc(signal, fs, n_mfcc=target_size[0])
if target_size:
mfcc = scipy.misc.imresize(mfcc, target_size)
mfcc = mfcc.reshape(mfcc.shape[0], mfcc.shape[1], 1)
return mfcc
def noise_augmentation():
da.noise_augmentation(x, noise_files_small)