Upload app.py with huggingface_hub

app.py

@@ -51,6 +51,21 @@ def _trim(waveform: torch.Tensor, sr: int, max_seconds: float):
     return waveform, False
 
 
+DAV_HOP = 512
+
+
+def _align_hop(waveform: torch.Tensor, sr: int):
+    """Trim the tail to a whole DAV hop.
+
+    The windowed refinement raises 'window overlap left uncovered latent frames' when the
+    total sample count is not a multiple of the 512-sample DAV hop (the final window start
+    is aligned down and the last partial frame is never covered). Trimming to a whole hop
+    removes the uncovered frame; at 44.1 kHz that is at most ~11 ms of audio.
+    """
+    frames = waveform.shape[-1] // DAV_HOP
+    return waveform[..., : frames * DAV_HOP]
+
+
 @spaces.GPU(duration=120)
 def refine(
     audio_path: str,
@@ -75,6 +90,7 @@ def refine(
         raise gr.Error("Please upload an audio file or pick an example first.")
     waveform, sr = _load_waveform(audio_path)
     waveform, trimmed = _trim(waveform, sr, MAX_SECONDS)
+    waveform = _align_hop(waveform, sr)
     steps = int(steps)
     if steps < 1:
         steps = 32