def augment(input_dir: Path, output_dir: Path):
seed(SEED)
if not input_dir.exists:
raise Exception('Input directory does not exist.')
if not output_dir.exists:
print("Making output directory {}".format(output_dir))
output_dir.mkdir(parents=True)
filelist = set([x for x in input_dir.glob('*.wav')])
print("{} input files found".format(len(filelist)))
while len(filelist) > 1:
print("{} files remaining...".format(len(filelist)))
f1 = choice(tuple(filelist))
filelist.remove(f1)
f2 = choice(tuple(filelist))
filelist.remove(f2)
# load audio files and apply a random amount of:
# gain reduction in steps of -3 db from [-15 db, 0 db]
# varispeed between [0.9, 1.1]
reduction = [0, -3, -6, -12, -15]
f1 = AudioFile(path=f1).varispeed(uniform(0.9,
1.1)).gain(choice(reduction))
f2 = AudioFile(path=f2).varispeed(uniform(0.9,
1.1)).gain(choice(reduction))
# mix two audio files - random amount of overlap from [0.5 to 1.5]
f1.mix(audio=f2, relative_start=uniform(0.5, 1.5))
# add a random amounts of silence [0 sec, 5 sec] before and after audio
f1.add_silence(sec_before=uniform(0, 5), sec_after=uniform(0, 5))
# save as new clean file
filename = f1.filename.split(".wav")[0] + "_" + f2.filename
f1.save(output_path=output_dir, filename=filename)
def augment(input_dir: Path, noise_dir: Path, output_dir: Path):
seed(SEED)
if not input_dir.exists:
raise Exception('Input directory does not exist.')
if not noise_dir.exists:
raise Exception('Noise directory does not exist.')
if not output_dir.exists:
print("Making output directory {}".format(output_dir))
output_dir.mkdir(parents=True)
filelist = [x for x in input_dir.glob('*.wav')]
print("{} input files found".format(len(filelist)))
noiselist = set([x for x in noise_dir.glob('*.wav')])
print("{} noise files found".format(len(noiselist)))
print("Loading noise files into memory...")
noise_files = [AudioFile(path=x) for x in noiselist]
print("Done loading noise files.")
irs = [x for x in Path('./IMreverbs-realistic').glob('*.wav')]
while len(filelist) > 0:
print("{} files remaining...".format(len(filelist)))
f1 = filelist.pop()
noise = choice(noise_files)
# load audio files and apply a random amount of processing to noisy file:
# gain reduction in steps of -6 db
# varispeed between [0.9, 1.1]
# start position of audio in noise file (trimming start of file)
gain = [6, 3, 1.5, 0, -1.5, -3, -6]
f1 = AudioFile(path=f1)
noise = noise.copy() \
.trim_start(relative_start=uniform(0.0, 0.5)) \
.trim_to_n_samples(n=f1.length)
# noise = noise.varispeed(uniform(0.9, 1.1))
noise = noise.gain(choice(gain)).clip()
# add dynamic lpf to simulate speaker turning away
filter_start = random()
filter_end = random()
if filter_start < filter_end:
f1.dynamic_lpf(cutoff=uniform(1000, 8000),
order=randint(0, 3),
relative_start=filter_start,
relative_end=filter_end,
exponential=random())
# add noise to audio
f1.mix(noise, maintain_length=True)
# choose random impulse response and add reverb to noisy audio
ir = AudioFile(path=choice(irs))
f1.conv_reverb(ir, wet_db=uniform(-70, -30), predelay=uniform(0, 50))
# filtering
f1.lpf(uniform(5000, 8000))
f1.hpf(uniform(0, 250))
# clipping
clipping = [0.0, 1.0, 2.0, 3.0]
f1.clip(choice(clipping))
# save
f1.save(output_path=output_dir)
