Initial working space

app.py

@@ -1,10 +1,16 @@
+import math
+
+import numpy as np
 import gradio
 import gradio.inputs
 import gradio.outputs
 import torch
-from df.enhance import enhance, init_df
+from df import config
+from df.enhance import enhance, init_df, load_audio, save_audio
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model, df, _ = init_df()
+model = model.to(device=device).eval()
 
 
 def mix_at_snr(clean, noise, snr, eps=1e-10):
@@ -21,22 +27,34 @@ def mix_at_snr(clean, noise, snr, eps=1e-10):
         mix: 1D Tensor with added clean and noise signals.
 
     """
-    clean = torch.as_tensor(clean)
-    noise = torch.as_tensor(noise)
+    clean = torch.as_tensor(clean).mean(0, keepdim=True)
+    noise = torch.as_tensor(noise).mean(0, keepdim=True)
+    if noise.shape[1] < clean.shape[1]:
+        noise = noise.repeat((1, int(math.ceil(clean.shape[1] / noise.shape[1]))))
+    noise = noise[:, : clean.shape[1]]
     E_speech = torch.mean(clean.pow(2)) + eps
     E_noise = torch.mean(noise.pow(2))
     K = torch.sqrt((E_noise / E_speech) * 10 ** (snr / 10) + eps)
     noise = noise / K
     mixture = clean + noise
-    assert torch.isfinite(mixture)
+    assert torch.isfinite(mixture).all()
     return clean, noise, mixture
 
+def as_gradio_audio(x):
+    sr = config.get("sr", "df", int)
+    return sr, (x/0x7fff).to(torch.int16).cpu().numpy()
 
 def mix_and_denoise(speech, noise, snr):
-    model, df, _ = init_df()
+    print(speech, noise, snr)
+    sr = config.get("sr", "df", int)
+    speech, _ = load_audio(speech, sr)
+    noise, _ = load_audio(noise, sr)
     speech, noise, noisy = mix_at_snr(speech, noise, snr)
-    enhanced = enhance(model.to(device=device).eval(), df, noisy)
-    return speech, noisy, enhanced
+    enhanced = enhance(model, df, noisy)
+    save_audio("clean.wav", speech, sr)
+    save_audio("noisy.wav", noisy, sr)
+    save_audio("enhanced.wav", enhanced, sr)
+    return "clean.wav", "noisy.wav", "enhanced.wav"
 
 
 inputs = [
@@ -49,8 +67,7 @@ inputs = [
     gradio.inputs.Slider(minimum=-10, maximum=40, step=5, default=10),
 ]
 examples = [
-    [],
-    ["samples/noise_freesound_2530.wav", "samples/noise_freesound_573577.wav"],
+    ["samples/p232_013_clean.wav", "samples/noise_freesound_2530.wav", 10],
 ]
 outputs = [
     gradio.outputs.Audio(label="Clean"),

samples/noise_freesound_573577.wav → clean.wav

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cb367b36e4e9d72d112377dd57bf354e13f0b30f8402c9e841ac47639e773497
-size 474010
+oid sha256:d7a51b4fdfb02657cf9410dbd34b4ea165acbec48581a8a074e1d45fdd3b3334
+size 378612

enhanced.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97ff9dc5db07e3a2410f0dd416d9bccdcdc9bd173ed46f415e405208a4105d04
+size 378284

noisy.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3b658209be05042ce017aa2b3db444e56e84c3cc6f58535599ff8887c9ee5f7
+size 378612

samples/p232_013_clean.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7a51b4fdfb02657cf9410dbd34b4ea165acbec48581a8a074e1d45fdd3b3334
+size 378612