More scripts for splitting and formatting audio

codes/scripts/audio/spleeter_split_voice_and_background.py

@@ -0,0 +1,62 @@
+from scipy.io import wavfile
+from spleeter.separator import Separator
+from tqdm import tqdm
+
+from data.util import find_audio_files
+import os
+import os.path as osp
+from spleeter.audio.adapter import AudioAdapter
+import numpy as np
+
+
+# Uses spleeter to divide audio clips into one of two bins:
+# 1. Audio has little to no background noise, saved to "output_dir"
+# 2. Audio has a lot of background noise, bg noise split off and saved to "output_dir_bg"
+if __name__ == '__main__':
+    src_dir = 'F:\\split\\books1'
+    output_dir = 'F:\\split\\cleaned\\books1'
+    output_dir_bg = 'F:\\split\\background-noise\\books1'
+    output_sample_rate=22050
+
+    os.makedirs(output_dir_bg, exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+
+    audio_loader = AudioAdapter.default()
+    separator = Separator('spleeter:2stems')
+    files = find_audio_files(src_dir, include_nonwav=True)
+    for e, file in enumerate(tqdm(files)):
+        file_basis = osp.relpath(file, src_dir)\
+            .replace('/', '_')\
+            .replace('\\', '_')\
+            .replace('.', '_')\
+            .replace(' ', '_')\
+            .replace('!', '_')\
+            .replace(',', '_')
+        if len(file_basis) > 100:
+            file_basis = file_basis[:100]
+        try:
+            wave, sample_rate = audio_loader.load(file, sample_rate=output_sample_rate)
+        except:
+            print(f"Error with {file}")
+            continue
+
+        sep = separator.separate(wave)
+        vocals = sep['vocals']
+        bg = sep['accompaniment']
+        vmax = np.abs(vocals).mean()
+        bmax = np.abs(bg).mean()
+
+        # Only output to the "good" sample dir if the ratio of background noise to vocal noise is high enough.
+        ratio = vmax / (bmax+.0000001)
+        if ratio >= 25:  # These values were derived empirically
+            od = output_dir
+            os = wave
+        elif ratio <= 1:
+            od = output_dir_bg
+            os = bg
+
+        # Strip out channels.
+        if len(os.shape) > 1:
+            os = os[:, 0]  # Just use the first channel.
+
+        wavfile.write(osp.join(od, f'{e}_{file_basis}.wav'), output_sample_rate, os)

codes/scripts/audio/split_on_silence.py

@@ -0,0 +1,42 @@
+import argparse
+import logging
+import os
+from pydub import AudioSegment
+from pydub.exceptions import CouldntDecodeError
+from pydub.silence import split_on_silence
+from data.util import find_audio_files
+from tqdm import tqdm
+
+
+# Uses pydub to process a directory of audio files, splitting them into clips at points where it detects a small amount
+# of silence.
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--path')
+    parser.add_argument('--out')
+    args = parser.parse_args()
+    minimum_duration = 5
+    maximum_duration = 20
+    files = find_audio_files(args.path, include_nonwav=True)
+    for e, wav_file in enumerate(tqdm(files)):
+        if e < 4197:
+            continue
+        print(f"Processing {wav_file}..")
+        outdir = os.path.join(args.out, f'{e}_{os.path.basename(wav_file[:-4])}').replace('.', '').strip()
+        os.makedirs(outdir, exist_ok=True)
+
+        try:
+            speech = AudioSegment.from_file(wav_file)
+        except CouldntDecodeError as e:
+            print(e)
+            continue
+        chunks = split_on_silence(speech, min_silence_len=300, silence_thresh=-40,
+                                  seek_step=100, keep_silence=50)
+
+        for i in range(0, len(chunks)):
+            if chunks[i].duration_seconds < minimum_duration or chunks[i].duration_seconds > maximum_duration:
+                continue
+            chunks[i].export(f"{outdir}/{i:05d}.wav", format='wav', parameters=["-ar", "22050", "-ac", "1"])
+
+if __name__ == '__main__':
+    main()

codes/scripts/audio/test_audio_gen.py

@@ -54,7 +54,7 @@ if __name__ == "__main__":
     torch.backends.cudnn.benchmark = True
     want_metrics = False
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to options YAML file.', default='../options/test_stop_pred_dataset.yml')
+    parser.add_argument('-opt', type=str, help='Path to options YAML file.', default='../options/test_lrdvae_audio_clips.yml')
     opt = option.parse(parser.parse_args().opt, is_train=False)
     opt = option.dict_to_nonedict(opt)
     utils.util.loaded_options = opt

codes/scripts/audio/test_audio_speech_recognition.py

@@ -44,7 +44,7 @@ if __name__ == "__main__":
     torch.backends.cudnn.benchmark = True
     want_metrics = False
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to options YAML file.', default='../options/test_gpt_asr_mozcv.yml')
+    parser.add_argument('-opt', type=str, help='Path to options YAML file.', default='../options/test_gpt_asr_mass.yml')
     opt = option.parse(parser.parse_args().opt, is_train=False)
     opt = option.dict_to_nonedict(opt)
     utils.util.loaded_options = opt