diff --git a/PyTorch/SpeechSynthesis/FastPitch/common/stft.py b/PyTorch/SpeechSynthesis/FastPitch/common/stft.py
index bc140c142..6a6cfc567 100644
--- a/PyTorch/SpeechSynthesis/FastPitch/common/stft.py
+++ b/PyTorch/SpeechSynthesis/FastPitch/common/stft.py
@@ -123,7 +123,7 @@ def inverse(self, magnitude, phase):
                 np.where(window_sum > tiny(window_sum))[0])
             window_sum = torch.autograd.Variable(
                 torch.from_numpy(window_sum), requires_grad=False)
-            window_sum = window_sum.cuda() if magnitude.is_cuda else window_sum
+            window_sum = window_sum.cuda(inverse_transform.get_device()) if magnitude.is_cuda else window_sum
             inverse_transform[:, :, approx_nonzero_indices] /= window_sum[approx_nonzero_indices]
 
             # scale by hop ratio