eschmidbauer/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Make spleeter filter simpler (and hopefully much faster)

Browse Source
This commit is contained in:
James Betker 2021-09-17 15:29:42 -06:00
parent 359e9e27a7
commit 97ea329a59
1 changed files with 5 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
codes/scripts/audio/preparation/spleeter_split_voice_and_background_2.py
View File

@ -12,13 +12,12 @@ from scripts.audio.preparation.spleeter_dataset import SpleeterDataset
def main(): def main():
src_dir = 'F:\\split\\joe_rogan' src_dir = 'F:\\split\\joe_rogan'
output_dir = 'F:\\split\\cleaned\\joe_rogan'
output_dir_bg = 'F:\\split\\background-noise\\joe_rogan'
output_sample_rate=22050 output_sample_rate=22050
batch_size=16 batch_size=16
dl = DataLoader(SpleeterDataset(src_dir, output_sample_rate, skip=batch_size*33000), batch_size=batch_size, shuffle=False, num_workers=1, pin_memory=True) dl = DataLoader(SpleeterDataset(src_dir, output_sample_rate, skip=batch_size*33000), batch_size=batch_size, shuffle=False, num_workers=1, pin_memory=True)
separator = Separator('pretrained_models/2stems', input_sr=output_sample_rate) separator = Separator('pretrained_models/2stems', input_sr=output_sample_rate)
unacceptable_files = open('unacceptable.txt', 'a')
for batch in tqdm(dl): for batch in tqdm(dl):
waves = batch['wave'] waves = batch['wave']
paths = batch['path'] paths = batch['path']
@ -26,7 +25,6 @@ def main():
sep = separator.separate(waves) sep = separator.separate(waves)
for j in range(sep['vocals'].shape[0]): for j in range(sep['vocals'].shape[0]):
wave = waves[j].cpu().numpy()[:durations[j]]
vocals = sep['vocals'][j][:durations[j]] vocals = sep['vocals'][j][:durations[j]]
bg = sep['accompaniment'][j][:durations[j]] bg = sep['accompaniment'][j][:durations[j]]
vmax = np.abs(vocals[output_sample_rate:-output_sample_rate]).mean() vmax = np.abs(vocals[output_sample_rate:-output_sample_rate]).mean()
@ -34,27 +32,10 @@ def main():
# Only output to the "good" sample dir if the ratio of background noise to vocal noise is high enough. # Only output to the "good" sample dir if the ratio of background noise to vocal noise is high enough.
ratio = vmax / (bmax+.0000001) ratio = vmax / (bmax+.0000001)
if ratio >= 4: # These values were derived empirically if ratio < 4: # These values were derived empirically
od = output_dir unacceptable_files.write(f'{paths[j]}\n')
out_sound = wave unacceptable_files.flush()
elif ratio <= 2: unacceptable_files.close()
od = output_dir_bg
out_sound = bg
else:
print(f"Reject {paths[j]}: {ratio}")
continue
# Strip out channels.
if len(out_sound.shape) > 1:
out_sound = out_sound[:, 0] # Just use the first channel.
# Compile an output path.
path = paths[j]
reld = str(os.path.relpath(os.path.dirname(path), src_dir)).strip()
os.makedirs(os.path.join(od, reld), exist_ok=True)
output_path = os.path.join(od, reld, os.path.basename(path))
wavfile.write(output_path, output_sample_rate, out_sound)
# Uses torch spleeter to divide audio clips into one of two bins: # Uses torch spleeter to divide audio clips into one of two bins: