Add filtration script for finding resampled clips (or phone calls)

codes/scripts/audio/preparation/filter_clips_with_no_hifreq_data.py

@@ -0,0 +1,23 @@
+import torch
+import torchaudio
+
+from data.audio.unsupervised_audio_dataset import load_audio
+from scripts.do_to_files import do_to_files
+
+
+def get_spec_mags(clip):
+    stft = torch.stft(clip, n_fft=22000, hop_length=1024, return_complex=True)
+    stft = stft[0, -2000:, :]
+    return (stft.real ** 2 + stft.imag ** 2).sqrt()
+
+
+def filter_no_hifreq_data(path, output_path):
+    clip = load_audio(path, 22050)
+    if clip.shape[-1] < 22050:
+        return
+    stft = get_spec_mags(clip)
+    if stft.mean() < .08:
+        print(f"Ignore {path}")
+
+if __name__ == '__main__':
+    do_to_files(filter_no_hifreq_data)

codes/scripts/do_to_files.py

@@ -0,0 +1,44 @@
+import argparse
+import functools
+import os
+import pathlib
+from multiprocessing.pool import ThreadPool
+
+from tqdm import tqdm
+
+
+''' 
+Helper function for scripts that iterate over large sets of files. Defines command-line arguments
+for operating over a large set of files, then handles setting up a worker queue system to operate
+on those files. You need to provide your own process_file_fn.
+
+process_file_fn expected signature:
+ (path, output_path)
+'''
+def do_to_files(process_file_fn):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--path')
+    parser.add_argument('--glob')
+    parser.add_argument('--out')
+    parser.add_argument('--resume')
+    parser.add_argument('--num_workers')
+
+    args = parser.parse_args()
+    src = args.path
+    glob = args.glob
+    out = args.out
+    resume = args.resume
+    num_workers = int(args.num_workers)
+
+    os.makedirs(out, exist_ok=True)
+    path = pathlib.Path(src)
+    files = path.rglob(glob)
+    files = [str(f) for f in files]
+    files = files[resume:]
+    pfn = functools.partial(process_file_fn, output_path=out)
+    if num_workers > 0:
+        with ThreadPool(num_workers) as pool:
+            list(tqdm(pool.imap(pfn, files), total=len(files)))
+    else:
+        for f in tqdm(files):
+            pfn(f)