Update app.py

app.py

@@ -1,23 +1,19 @@
-import torch
-import torchaudio
-from pydub import AudioSegment
-import gradio as gr
-_original_load = torch.load
-def cpu_load(*args, **kwargs):
-    kwargs["map_location"] = torch.device('cpu')
-    return _original_load(*args, **kwargs)
-torch.load = cpu_load
-
-hubert = torch.hub.load("bshall/hubert:main", "hubert_soft", trust_repo=True).cpu()
-acoustic = torch.hub.load("bshall/acoustic-model:main", "hubert_soft", trust_repo=True).cpu()
-hifigan = torch.hub.load("bshall/hifigan:main", "hifigan_hubert_soft", trust_repo=True).cpu()
-
-def soft_vc(audio_path):
-  AudioSegment.from_file(audio_path).set_frame_rate(16000).set_channels(1).export(audio_path, format="wav")
-  source = torchaudio.load(audio_path)[0].unsqueeze(0).cpu()
-  with torch.inference_mode():
-      target = hifigan(acoustic.generate(hubert.units(source)).transpose(1, 2))
-  torchaudio.save("output.wav", target.squeeze(0).cpu(), 16000)
-  return "output.wav"
-
+import torch
+import torchaudio
+import gradio as gr
+from torchaudio.functional import resample
+hubert = torch.hub.load("bshall/hubert:main", "hubert_soft", trust_repo=True)
+acoustic = torch.hub.load("bshall/acoustic-model:main", "hubert_soft", trust_repo=True)
+hifigan = torch.hub.load("bshall/hifigan:main", "hifigan_hubert_soft", trust_repo=True)
+
+@torch.inference_mode()
+def soft_vc(audio_path):
+    wav,sr = torchaudio.load(audio_path)
+    if sr != 16000:
+        resample(wav,sr,16000)
+    wav = wav.mean(0,True)
+    torchaudio.save("output.wav", hifigan(acoustic.generate(hubert.units(wav.unsqueeze(0))).transpose(1, 2)).squeeze(0), 16000)
+    return "output.wav"
+
+
 gr.Interface(soft_vc,gr.Audio(label="Input Audio",type="filepath"),gr.Audio(label="Output Audio",type="filepath")).launch()