Data Augmentation

Data Augmentation

Same Class Augmentation

The same class augmentation simply takes two signal segments s1, and s2, and additively combines them according to:

s_aug = alpha * s1 + (1-alpha) * s2

where alpha is drawn at random from the uniform distribution [0, 1).

Same Class Augmentation

Python Implementation

def same_class_augmentation(wave, class_dir):
    """ Perform same class augmentation of the wave by loading a random segment
    from the class_dir and additively combine the wave with that segment.
    """
    sig_paths = glob.glob(os.path.join(class_dir, "*.wav"))
    aug_sig_path = np.random.choice(sig_paths, 1, replace=False)[0]
    (fs, aug_sig) = utils.read_wave_file(aug_sig_path)
    alpha = np.random.rand()
    wave = (1.0-alpha)*wave + alpha*aug_sig
    return wave

Noise Augmentation

Noise augmentation is simply an addition of a random noise segment on top of the augmented signal with a dampening factor of 0.4.

s_aug += (noise_seg * 0.4) for noise_seg in noise_segs

where noise_segs are three randomly chosen noise segments from the directory of noise segments.

Noise Augmentation

Python Implementation

def noise_augmentation(wave, noise_dir):
    """ Perform noise augmentation of the wave by loading three noise segments
    from the noise_dir and add these on top of the wave with a dampening factor
    of 0.4
    """
    noise_paths = glob.glob(os.path.join(noise_dir, "*.wav"))
    aug_noise_paths = np.random.choice(noise_paths, 3, replace=False)
    dampening_factor = 0.4
    for aug_noise_path in aug_noise_paths:
        (fs, aug_noise) = utils.read_wave_file(aug_noise_path)
        wave = wave + aug_noise*dampening_factor
    return wave

Time Shift Augmentation

The time shift augmentation is a random shift of the signal by rolling the signal along the time axis. A wrap around shift.

def time_shift_spectrogram(spectrogram):
    """ Shift a spectrogram along the time axis in the spectral-domain at random
    """
    nb_cols = spectrogram.shape[1]
    nb_shifts = np.random.randint(0, nb_cols)

    return np.roll(spectrogram, nb_shifts, axis=1)

Pitch Shift Augmentation

The pitch shift augmentation is a random roll in the range +-5% around the frequency axis. A wrap-around shift to preserve all information.

def pitch_shift_spectrogram(spectrogram):
    """ Shift a spectrogram along the frequency axis in the spectral-domain at
    random
    """
    nb_cols = spectrogram.shape[0]
    max_shifts = nb_cols//20 # around 5% shift
    nb_shifts = np.random.randint(-max_shifts, max_shifts)

    return np.roll(spectrogram, nb_shifts, axis=0)