Hendrik Schroeter commited on
Commit
df38a50
1 Parent(s): 973537b

add random noise start

Browse files
Files changed (1) hide show
  1. app.py +4 -4
app.py CHANGED
@@ -30,7 +30,9 @@ def mix_at_snr(clean, noise, snr, eps=1e-10):
30
  noise = torch.as_tensor(noise).mean(0, keepdim=True)
31
  if noise.shape[1] < clean.shape[1]:
32
  noise = noise.repeat((1, int(math.ceil(clean.shape[1] / noise.shape[1]))))
33
- noise = noise[:, : clean.shape[1]]
 
 
34
  E_speech = torch.mean(clean.pow(2)) + eps
35
  E_noise = torch.mean(noise.pow(2))
36
  K = torch.sqrt((E_noise / E_speech) * 10 ** (snr / 10) + eps)
@@ -64,9 +66,7 @@ inputs = [
64
  gradio.inputs.Audio(
65
  source="microphone,upload", type="filepath", optional=True, label="Speech"
66
  ),
67
- gradio.inputs.Audio(
68
- source="upload", type="filepath", optional=True, label="Noise"
69
- ),
70
  gradio.inputs.Slider(minimum=-10, maximum=40, step=5, default=10),
71
  ]
72
  examples = [
 
30
  noise = torch.as_tensor(noise).mean(0, keepdim=True)
31
  if noise.shape[1] < clean.shape[1]:
32
  noise = noise.repeat((1, int(math.ceil(clean.shape[1] / noise.shape[1]))))
33
+ max_start = int(noise.shape[1] - clean.shape[1])
34
+ start = torch.randint(0, max_start)
35
+ noise = noise[:, start : start + clean.shape[1]]
36
  E_speech = torch.mean(clean.pow(2)) + eps
37
  E_noise = torch.mean(noise.pow(2))
38
  K = torch.sqrt((E_noise / E_speech) * 10 ** (snr / 10) + eps)
 
66
  gradio.inputs.Audio(
67
  source="microphone,upload", type="filepath", optional=True, label="Speech"
68
  ),
69
+ gradio.inputs.Audio(source="upload", type="filepath", optional=True, label="Noise"),
 
 
70
  gradio.inputs.Slider(minimum=-10, maximum=40, step=5, default=10),
71
  ]
72
  examples = [