fixed denoiser="demucs" + suppress_silence=False error (#441)

-fixed tensor to numpy conversion error caused by using `denoiser="demucs"` with `suppress_silence=False` when device is not CPU (#441)

-fixed incorrect return type hint for `.stabilization.nonvad.wav2mask()`

Author: jianfch

stable_whisper/stabilization/__init__.py

@@ -98,7 +98,8 @@ def reset(self):
     def _silent_mask_test(self, mask, min_unit_per_word) -> bool:
         if self.ignore_is_silent or mask is None:
             return False
-        non_silent_unit_count = mask.shape[-1] - np.flatnonzero(mask).shape[-1]
+        nonzero_count = torch.count_nonzero(mask) if torch.is_tensor(mask) else np.count_nonzero(mask)
+        non_silent_unit_count = mask.shape[-1] - nonzero_count
         return non_silent_unit_count < min_unit_per_word
 
     def _append_timings(self, timings: np.ndarray):

stable_whisper/stabilization/nonvad.py

@@ -45,7 +45,7 @@ def wav2mask(
         q_levels: int = 20,
         k_size: int = 5,
         sr: int = None
-) -> (Tuple[torch.Tensor, Tuple[np.ndarray, np.ndarray]], None):
+) -> (torch.Tensor, None):
     """
     Generate 1D mask from waveform for suppressing timestamp tokens.
     """