import librosa
import random
from smd import config
from scipy.ndimage import affine_transform, spline_filter
import numpy as np
import warnings
def pitch_shifting_audio(audio, n_steps=None):
if n_steps is None:
n_steps = float(random.randint(int(config.SHIFTING_MIN * 2), int(config.SHIFTING_MAX * 2))) / 2
return librosa.effects.pitch_shift(audio, config.SAMPLING_RATE, n_steps=n_steps)
def time_stretching_audio(audio, rate=None):
if rate is None:
rate = random.uniform(config.STRETCHING_MIN, config.STRETCHING_MAX)
return librosa.effects.time_stretch(audio, rate), rate
def pitch_time_deformation_spec(spec, stretch_rate=None, shift_rate=None):
""" Based on https://github.com/f0k/ismir2018/blob/master/experiments/augment.py """
shape = spec.shape
random = (np.random.rand(2) - .5) * 2
warnings.filterwarnings("ignore", "The behaviour of affine_transform with a "
"one-dimensional array supplied for the matrix parameter "
"has changed in scipy 0.18.0.", module="scipy.ndimage")
if stretch_rate is None:
stretch_rate = 1 + random[0] * config.MAX_STRETCHING
if shift_rate is None:
shift_rate = 1 + random[1] * config.MAX_SHIFTING_PITCH
new_length = int(shape[1] * stretch_rate)
# We can do shifting/stretching and cropping in a single affine
# transform (including setting the upper bands to zero if we shift
# down the signal so far that we exceed the input's nyquist rate)
spec = spline_filter(spec.T, 2).astype(spec.dtype)
spec = affine_transform(spec, (1 / stretch_rate, 1 / shift_rate),
output_shape=(new_length, shape[0]),
offset=(0, 0), mode='constant', order=2,
prefilter=False)
# clip possible negative values introduced by the interpolation
# delete the last frame if it is empty
if (spec[-1] == 0).all():
spec = spec[:-1]
return np.maximum(spec.T, 0), stretch_rate
