def create_original_spectrograms(self, samples, n, species_wav_input_dir,
species_spectrogram_output_dir):
samples = random.sample(samples, int(n)) # choose n from all samples
for sample_name in samples:
SoundProcessor.create_spectrogram(sample_name,
species_wav_input_dir,
species_spectrogram_output_dir,
n_mels=128)
def test_create_one_spectro(self):
# Test creating one spectrogram from an audio
# file. Ensure that sample rate, duration, and
# species are included int the destination spectrogram
# .png file:
audio_path = os.path.join(self.cur_dir,
'audio_aug_tst_data/DYSMEN_S/dys1.mp3')
(aud, sr) = SoundProcessor.load_audio(audio_path)
with tempfile.NamedTemporaryFile(suffix='.png',
prefix='spectro',
dir='/tmp',
delete=True) as fd:
SoundProcessor.create_spectrogram(aud,
sr,
fd.name,
info={'species': 'DYSMEN_C'})
_spectro, info = SoundProcessor.load_spectrogram(fd.name)
truth = {'sr': '22050', 'duration': '10.8', 'species': 'DYSMEN_C'}
self.assertDictEqual(info, truth)
def create_new_sample(self, sample_name, paths, num_augs=1):
(species_wav_input_dir, species_wav_output_dir,
species_spectrogram_output_dir) = paths
aug_choices = np.random.choice(self.AUDIO_AUG_NAMES,
size=num_augs,
p=self.P_DIST,
replace=False)
# input(f"Aug choices: {aug_choices}")
# Warping must be done after all the other augmentations take place,
# after spectrogram is created
warp = False
if "warp" in aug_choices:
warp = True
aug_choices = aug_choices.tolist()
# print(f"Aug chioces as list: {aug_choices}")
aug_choices.remove("warp")
# print(f"Aug chioces after: {aug_choices}")
for i in range(len(aug_choices)):
# print(aug_choices)
aug_name = aug_choices[i]
if i != 0: # if not first augmentation, then, source wav is in output wav directory
species_wav_input_dir = species_wav_output_dir
if aug_name == "add_noise":
# Add_noise; which noise to add will be chosen at random
updated_name = SoundProcessor.add_background(
sample_name,
self.NOISE_PATH,
species_wav_input_dir,
species_wav_output_dir,
len_noise_to_add=5.0)
elif aug_name == "time_shift":
updated_name = SoundProcessor.time_shift(
sample_name, species_wav_input_dir, species_wav_output_dir)
sample_name = updated_name
# create new spectrogram if augmented
if len(aug_choices) != 0:
sample_name = SoundProcessor.create_spectrogram(
sample_name,
species_wav_output_dir,
species_spectrogram_output_dir,
n_mels=128)
if warp:
#warp
# if len(aug_choices) +1 > 1:
# input(f"num_augs = {len(aug_choices) +1} for {sample_name}")
sample_name = sample_name[:-len(".wav")] + ".png"
# Above: if sample is unaugmented to this point, sample_name will be
# *.wav. Since SoundProcessor.warp_spectrogram expects sample_name to be *.png, we
# replace extension. If augmented and sample_name is already *.png,
# there is no change.
warped_name = SoundProcessor.warp_spectrogram(
sample_name, species_spectrogram_output_dir,
species_spectrogram_output_dir)
# if warp is not the only augmentation,
# we do not want spectrogram before warp
if len(aug_choices) != 0:
assert (warped_name != sample_name)
fname = os.path.join(species_spectrogram_output_dir,
sample_name)
os.remove(fname)
def chop_one_audio_file(self, in_dir, species, spectro_fname, out_dir, window_len = 5):
"""
Generates window_len second sound file snippets
and associated spectrograms from sound files of
arbitrary length.
Performs a time shift on all the wav files in the
species directories. The shift is 'rolling' such that
no information is lost.
:param in_dir: directory of the audio file to chop
:type file_name: str
:param species: the directory names of the species to
modify the wav files of. If species=None, all
subdirectories will be processed.
:type species: {None | [str]}
:param spectro_fname: basefile name of audio file to chop
:type spectro_fname: str
:param out_dir: root directory under which spectrogram
and audio snippets will be saved (in different subdirs)
:type out_dir: str
"""
orig, sample_rate = librosa.load(os.path.join(in_dir, species, spectro_fname))
length = int(librosa.get_duration(orig, sample_rate))
for start_time in range(length - window_len):
fpath = Path(spectro_fname)
window_name = f"{fpath.stem}_sw-start{str(start_time)}"
window_file_name = str(Path.joinpath(fpath.parent, window_name))
outfile_spectro = os.path.join(out_dir,
'spectrograms/',
species,
f"{window_file_name}.png")
outfile_audio = os.path.join(out_dir,
'wav-files',
species,
f"{window_file_name}.{'wav'}")
spectro_done = os.path.exists(outfile_spectro)
audio_done = os.path.exists(outfile_audio)
if spectro_done and audio_done and WhenAlreadyDone.SKIP:
# No brainer no need to even read the audio excerpt:
continue
if spectro_done and not audio_done and not self.generate_wav_files:
continue
# Need an audio snippet either for
# a spectrogram or wav file:
window_audio, sr = librosa.load(os.path.join(in_dir, species, spectro_fname),
offset=start_time, duration=window_len)
if not spectro_done or (spectro_done and self.overwrite_policy != WhenAlreadyDone.SKIP):
SoundProcessor.create_spectrogram(window_audio,sr,outfile_spectro)
if self.generate_wav_files:
if audio_done and self.overwrite_policy == WhenAlreadyDone.SKIP:
continue
else:
sf.write(outfile_audio, window_audio, sr)