Patch v2Pro speaker preprocessing for torchaudio

app.py

@@ -373,6 +373,33 @@ def patch_upstream_repo():
     if old in content:
         content = content.replace(old, new, 1)
         chinese2.write_text(content, encoding="utf-8")
+    sv_script = GPT_SOVITS_DIR / "GPT_SoVITS" / "prepare_datasets" / "2-get-sv.py"
+    sv_content = sv_script.read_text(encoding="utf-8")
+    if "from scipy.io import wavfile" not in sv_content:
+        sv_content = sv_content.replace(
+            "import torch\n",
+            "import torch\nimport numpy as np\nfrom scipy.io import wavfile\n",
+            1,
+        )
+    old_load = """    wav32k, sr0 = torchaudio.load(wav_path)
+    assert sr0 == 32000
+    wav32k = wav32k.to(device)
+"""
+    new_load = """    sr0, wav32k_np = wavfile.read(wav_path)
+    assert sr0 == 32000
+    if wav32k_np.ndim == 1:
+        wav32k_np = wav32k_np[None, :]
+    else:
+        wav32k_np = wav32k_np.T
+    if np.issubdtype(wav32k_np.dtype, np.integer):
+        wav32k_np = wav32k_np.astype("float32") / float(np.iinfo(wav32k_np.dtype).max)
+    else:
+        wav32k_np = wav32k_np.astype("float32")
+    wav32k = torch.from_numpy(wav32k_np).to(device)
+"""
+    if old_load in sv_content:
+        sv_content = sv_content.replace(old_load, new_load, 1)
+        sv_script.write_text(sv_content, encoding="utf-8")
     patch_marker.write_text("ok\n", encoding="utf-8")