Extract some preprocessing from private repo.

src/torch_wae/audio.py

@@ -0,0 +1,48 @@
+from __future__ import annotations
+
+from random import Random
+
+import torch
+
+
+def omit_silence(
+    waveform: torch.Tensor,
+    threshold: float,
+    channel_first: bool = False,
+) -> torch.Tensor:
+    abs_waveform = torch.abs(waveform)
+    not_silent = abs_waveform > threshold
+    indices_not_silent = torch.where(
+        torch.any(not_silent, dim=1 if channel_first else 0)
+    )[0]
+
+    if len(indices_not_silent) == 0:
+        waveform = torch.zeros_like(waveform)
+    else:
+        begin_non_silence = indices_not_silent[0]
+        end_non_silence = indices_not_silent[-1]
+        if channel_first:
+            waveform = waveform[begin_non_silence : end_non_silence + 1, :]
+        else:
+            waveform = waveform[:, begin_non_silence : end_non_silence + 1]
+
+    return waveform
+
+
+def crop_randomly(
+    x: torch.Tensor,
+    random: Random,
+    sample_rate: int,
+    durations: int,
+) -> torch.Tensor:
+    c, d = x.shape
+    size = sample_rate * durations
+    pad = max(0, size - d)
+
+    if pad == 0:
+        start = random.randint(0, d - size)
+        end = start + size
+        return x[:, start:end]
+    else:
+        p = torch.zeros((c, pad), dtype=x.dtype).to(x.device)
+        return torch.cat((x, p), dim=-1)