def compose_without_noise(ir_path='data/impulse'):
_p = 0.25
transforms = [
AddGaussianNoise(p=_p),
Shift(p=_p, min_fraction=-0.2, max_fraction=0.2),
FrequencyMask(p=_p),
TimeMask(p=_p, max_band_part=0.25),
AddGaussianSNR(p=_p),
ClippingDistortion(p=_p, max_percentile_threshold=20),
MyAddImpulseResponse(p=_p, ir_path=ir_path),
TimeStretch(p=_p / 10),
PitchShift(p=_p / 25),
]
return MyCompose(transforms, p=1.0, max_augs=3)
def compose(sounds_path):
_p = 0.2
transforms = [
MyGain(p=_p),
AddGaussianNoise(p=_p),
Shift(p=_p, min_fraction=-0.25, max_fraction=0.25),
FrequencyMask(p=_p),
TimeMask(p=_p, max_band_part=0.25),
AddGaussianSNR(p=_p),
ClippingDistortion(p=_p, max_percentile_threshold=20),
AddBackgroundNoise(sounds_path=sounds_path, p=_p),
TimeStretch(p=_p/10),
PitchShift(p=_p/30),
]
return Compose(transforms, p=0.4, shuffle=True)
min_time_between_sounds=2.0,
max_time_between_sounds=4.0,
burst_probability=0.4,
min_pause_factor_during_burst=0.01,
max_pause_factor_during_burst=0.95,
min_fade_in_time=0.005,
max_fade_in_time=0.08,
min_fade_out_time=0.01,
max_fade_out_time=0.1,
p=1.0,
),
"num_runs":
5,
},
{
"instance": ClippingDistortion(p=1.0),
"num_runs": 5
},
{
"instance":
FrequencyMask(min_frequency_band=0.5,
max_frequency_band=0.6,
p=1.0),
"num_runs":
5,
},
{
"instance": Gain(min_gain_in_db=-6, max_gain_in_db=6, p=1.0),
"num_runs": 5
},
{
def generate(self, wave_file, output_dir):
"""
For each transformation, apply it to an example sound and write the transformed sounds to
an output folder.
"""
samples = load_wav_file(wave_file)
_filename = os.path.basename(wave_file).split('.')[0]
# AddImpulseResponse
if self.AddImpulseResponse[0]:
augmenter = Compose([
AddImpulseResponse(p=1.0, ir_path=os.path.join(DEMO_DIR, "ir"))
])
output_file_path = os.path.join(
output_dir,
_filename + "_AddImpulseResponse{:03d}.wav".format(0))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# FrequencyMask
if self.FrequencyMask[0]:
augmenter = Compose([FrequencyMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_FrequencyMask{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeMask
if self.TimeMask[0]:
augmenter = Compose([TimeMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_TimeMask{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddGaussianSNR
if self.AddGaussianSNR[0]:
augmenter = Compose([AddGaussianSNR(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddGaussianSNR{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddGaussianNoise
if self.AddGaussianNoise[0]:
augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001,
max_amplitude=0.015,
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddGaussianNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeStretch
if self.TimeStretch[0]:
augmenter = Compose(
[TimeStretch(min_rate=0.5, max_rate=1.5, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_TimeStretch{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# PitchShift
if self.PitchShift[0]:
augmenter = Compose(
[PitchShift(min_semitones=-6, max_semitones=12, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_PitchShift{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Shift
if self.Shift[0]:
augmenter = Compose(
[Shift(min_fraction=-0.5, max_fraction=0.5, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_Shift{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Shift without rollover
if self.ShiftWithoutRoll[0]:
augmenter = Compose([
Shift(min_fraction=-0.2,
max_fraction=0.2,
rollover=False,
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_ShiftWithoutRollover{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Normalize
if self.Normalize[0]:
augmenter = Compose([Normalize(p=1.0)])
output_file_path = os.path.join(
output_dir, _filename + "_Normalize{:03d}.wav".format(0))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Resample
if self.Resample[0]:
augmenter = Compose([
Resample(min_sample_rate=12000, max_sample_rate=44100, p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_Resample{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# ClippingDistortion
if self.ClippingDistortion[0]:
augmenter = Compose(
[ClippingDistortion(max_percentile_threshold=10, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_ClippingDistortion{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddBackgroundNoise
if self.AddBackgroundNoise[0]:
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
DEMO_DIR, "background_noises"),
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddBackgroundNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddWhiteNoise
if self.AddWhiteNoise[0]:
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
DEMO_DIR, "white_noises"),
p=1.0)
])
for i in range(self.AddWhiteNoise[1]):
output_file_path = os.path.join(
output_dir,
_filename + "_AddWhiteNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddPinkNoise
if self.AddPinkNoise[0]:
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
DEMO_DIR, "pink_noises"),
p=1.0)
])
for i in range(self.AddPinkNoise[1]):
output_file_path = os.path.join(
output_dir,
_filename + "_AddPinkNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddShortNoises
if self.AddShortNoises[0]:
augmenter = Compose([
AddShortNoises(
sounds_path=os.path.join(DEMO_DIR, "short_noises"),
min_snr_in_db=0,
max_snr_in_db=8,
min_time_between_sounds=2.0,
max_time_between_sounds=4.0,
burst_probability=0.4,
min_pause_factor_during_burst=0.01,
max_pause_factor_during_burst=0.95,
min_fade_in_time=0.005,
max_fade_in_time=0.08,
min_fade_out_time=0.01,
max_fade_out_time=0.1,
p=1.0,
)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddShortNoises{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
assert sound_np.dtype == np.int16
sound_np = np.divide(
sound_np, 32768, dtype=np.float32
)
number = os.path.split(audio_file)[-1][:-4]
transforms = [
{"instance": AddGaussianSNR(p=1.0), "num_runs": 3},
{"instance": TimeStretch(min_rate=0.4, max_rate=1.25, p=1.0), "num_runs": 5},
{
"instance": PitchShift(min_semitones=-5, max_semitones=5, p=1.0),
"num_runs": 6,
},
{"instance": Shift(min_fraction=-0.85, max_fraction=0.85, p=1.0), "num_runs": 4},
{"instance": Resample(p=1.0), "num_runs": 5},
{"instance": ClippingDistortion(p=1.0), "num_runs": 3},
]
for transform in transforms:
augmenter = Compose([transform["instance"]])
run_name = (
transform.get("name")
if transform.get("name")
else transform["instance"].__class__.__name__
)
for i in range(transform["num_runs"]):
output_file_path = os.path.join(
'augmented', "{}_{}_{:03d}.wav".format(number, run_name, i)
)
augmented_samples = augmenter(samples=sound_np, sample_rate=sample_rate)
wavfile.write(output_file_path, rate=sample_rate, data=augmented_samples)
def __init__(self, augment_type, p, cross_valid=False):
self.cross_valid = cross_valid
self.sample_rate = 8000
self.type = augment_type
self.p = p
wham_path = '../../../librimix/data/wham_noise/cv' if self.cross_valid else '../../../librimix/data/wham_noise/tr'
if self.type == 'wham_weak':
self.augment = Compose([
AddBackgroundNoise(sounds_path=wham_path,
min_snr_in_db=5,
max_snr_in_db=15,
p=1)
])
elif self.type == 'wham_strong':
self.augment = Compose([
AddBackgroundNoise(sounds_path=wham_path,
min_snr_in_db=2,
max_snr_in_db=7,
p=1)
])
elif self.type == 'reverb_weak':
self.augment = AudioEffectsChain().reverb(
reverberance=random.randrange(0, 50),
room_scale=random.randrange(0, 50),
stereo_depth=random.randrange(0, 50),
)
elif self.type == 'reverb_strong':
self.augment = AudioEffectsChain().reverb(
reverberance=random.randrange(50, 100),
room_scale=random.randrange(50, 100),
stereo_depth=random.randrange(50, 100),
)
elif self.type == 'cascade':
self.augment = Compose([
AddBackgroundNoise(sounds_path=wham_path,
min_snr_in_db=0,
max_snr_in_db=5,
p=self.p),
AddGaussianSNR(min_SNR=0.001, max_SNR=0.25, p=self.p),
ClippingDistortion(min_percentile_threshold=0,
max_percentile_threshold=40,
p=self.p),
FrequencyMask(min_frequency_band=0.0,
max_frequency_band=0.5,
p=self.p),
PolarityInversion(p=self.p),
Shift(min_fraction=-0.5,
max_fraction=0.5,
rollover=True,
p=self.p),
TimeMask(min_band_part=0.0,
max_band_part=0.2,
fade=False,
p=self.p)
])
elif self.type == 'distort':
self.augment = Compose([
PitchShift(min_semitones=-4, max_semitones=4, p=self.p),
TimeStretch(min_rate=0.8,
max_rate=1.25,
leave_length_unchanged=True,
p=self.p)
])
elif self.type == 'none':
self.augment = None
else:
raise ValueError(
"Did not recognize augmentation type. Received %s, expected 'wham_weak', 'wham_strong', 'reverb_weak', 'reverb_strong', 'cascade', 'distort', or 'none'."
% self.type)
def transform(file_path, output_folder, iterations):
"""
For each transformation, apply it to an example sound and write the transformed sounds to
an output folder.
"""
samples = load_wav_file(file_path)
file_name = os.path.basename(file_path).replace('.wav', '')
def produce(augmenter, name):
for i in range(iterations):
output_file_path = '{}/{}'.format(
output_folder, "{}_{}_{}.wav".format(name, file_name, i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeMask
augmenter = Compose([TimeMask(p=1.0)])
produce(augmenter, 'TimeMask')
# FrequencyMask
augmenter = Compose([FrequencyMask(p=1.0)])
produce(augmenter, 'FrequencyMask')
# AddGaussianSNR
augmenter = Compose([AddGaussianSNR(p=1.0)])
produce(augmenter, 'AddGaussianSNR')
# PitchShift
augmenter = Compose([PitchShift(min_semitones=-4, max_semitones=4, p=1.0)])
produce(augmenter, 'PitchShift')
# TimeStretch
augmenter = Compose([TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5)])
produce(augmenter, 'TimeStretch')
# AddGaussianNoise
augmenter = Compose(
[AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=1.0)])
produce(augmenter, 'AddGaussianNoise')
# Shift
augmenter = Compose([Shift(min_fraction=-0.5, max_fraction=0.5, p=1.0)])
produce(augmenter, 'Shift')
# Shift without rollover
augmenter = Compose(
[Shift(min_fraction=-0.5, max_fraction=0.5, rollover=False, p=1.0)])
produce(augmenter, 'Shift without rollover')
# Normalize
augmenter = Compose([Normalize(p=1.0)])
produce(augmenter, 'Normalize')
# AddImpulseResponse
augmenter = Compose(
[AddImpulseResponse(p=1.0, ir_path=os.path.join(DEMO_DIR, "ir"))])
produce(augmenter, 'AddImpulseResponse')
# Resample
augmenter = Compose([Resample(p=1.0)])
produce(augmenter, 'Resample')
# ClippingDistortion
augmenter = Compose([ClippingDistortion(p=1.0)])
produce(augmenter, 'ClippingDistortion')
# AddBackgroundNoise
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(DEMO_DIR,
"background_noises"),
p=1.0)
])
produce(augmenter, 'AddBackgroundNoise')
# AddShortNoises
augmenter = Compose([
AddShortNoises(
sounds_path=os.path.join(DEMO_DIR, "short_noises"),
min_snr_in_db=0,
max_snr_in_db=8,
min_time_between_sounds=2.0,
max_time_between_sounds=4.0,
burst_probability=0.4,
min_pause_factor_during_burst=0.01,
max_pause_factor_during_burst=0.95,
min_fade_in_time=0.005,
max_fade_in_time=0.08,
min_fade_out_time=0.01,
max_fade_out_time=0.1,
p=1.0,
)
])
produce(augmenter, 'AddShortNoises')
"Normalize_{:03d}.wav".format(0))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
# Resample
augmenter = Compose([Resample(p=1.0)])
for i in range(5):
output_file_path = os.path.join(output_dir,
"Resample_{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# ClippingDistortion
augmenter = Compose([ClippingDistortion(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "ClippingDistortion_{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddBackgroundNoise
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(DEMO_DIR,
"background_noises"),
p=1.0)
])
for i in range(5):
def applyTransformations(fileName, output_dir, auxiliarSoundsDir):
name = fileName.split(".")[0].split("/")[-1]
samples = load_wav_file(fileName)
# AddImpulseResponse
augmenter = Compose([
AddImpulseResponse(p=1.0,
ir_path=os.path.join(auxiliarSoundsDir,
"helperSounds/ir"))
])
output_file_path = os.path.join(
output_dir, "{}_AddImpulseResponse_{:03d}.wav".format(name, 0))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
# FrequencyMask
augmenter = Compose([FrequencyMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_FrequencyMask_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeMask
augmenter = Compose([TimeMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_TimeMask_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddGaussianSNR
augmenter = Compose([AddGaussianSNR(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_AddGaussianSNR_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddGaussianNoise
augmenter = Compose(
[AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_AddGaussianNoise_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeStretch
augmenter = Compose([TimeStretch(min_rate=0.8, max_rate=1.25, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_TimeStretch_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# PitchShift
augmenter = Compose([PitchShift(min_semitones=-4, max_semitones=4, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_itchShift_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Shift
augmenter = Compose([Shift(min_fraction=-0.5, max_fraction=0.5, p=1.0)])
for i in range(5):
output_file_path = os.path.join(output_dir,
"{}_Shift_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Shift without rollover
augmenter = Compose(
[Shift(min_fraction=-0.5, max_fraction=0.5, rollover=False, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_ShiftWithoutRollover_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Normalize
augmenter = Compose([Normalize(p=1.0)])
output_file_path = os.path.join(output_dir,
"{}_Normalize_{:03d}.wav".format(name, 0))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
# ClippingDistortion
augmenter = Compose([ClippingDistortion(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_ClippingDistortion_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddBackgroundNoise
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
auxiliarSoundsDir, "helperSounds/background_noises"),
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_AddBackgroundNoise_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddShortNoises
augmenter = Compose([
AddShortNoises(
sounds_path=os.path.join(auxiliarSoundsDir,
"helperSounds/short_noises"),
min_snr_in_db=0,
max_snr_in_db=8,
min_time_between_sounds=2.0,
max_time_between_sounds=4.0,
burst_probability=0.4,
min_pause_factor_during_burst=0.01,
max_pause_factor_during_burst=0.95,
min_fade_in_time=0.005,
max_fade_in_time=0.08,
min_fade_out_time=0.01,
max_fade_out_time=0.1,
p=1.0,
)
])
for i in range(5):
output_file_path = os.path.join(
output_dir, "{}_AddShortNoises_{:03d}.wav".format(name, i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)