.

Author: XXXXRT666

.github/build_windows_packages.ps1

@@ -134,6 +134,7 @@ switch ($cuda) {
 }
 
 Write-Host "[INFO] Installing dependencies..."
+& ".\runtime\python.exe" -m pip install --pre torchcodec --index-url https://download.pytorch.org/whl/nightly/cpu
 & ".\runtime\python.exe" -m pip install -r extra-req.txt --no-deps --no-warn-script-location
 & ".\runtime\python.exe" -m pip install -r requirements.txt --no-warn-script-location
 

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -721,20 +721,18 @@ def _set_ref_spec(self, ref_audio_path):
             self.prompt_cache["refer_spec"][0] = spec_audio
 
     def _get_ref_spec(self, ref_audio_path):
-        audio_n, raw_sr = librosa.load(ref_audio_path, sr=self.configs.sampling_rate)
-        raw_audio = torch.from_numpy(audio_n).unsqueeze(0)
-        raw_audio = raw_audio.to(self.configs.device).float()
+        raw_audio, raw_sr = torchaudio.load_with_torchcodec(ref_audio_path)
         self.prompt_cache["raw_audio"] = raw_audio
         self.prompt_cache["raw_sr"] = raw_sr
 
         if raw_sr != self.configs.sampling_rate:
             audio = raw_audio.to(self.configs.device)
-            if audio.shape[0] == 2:
+            if audio.shape[0] > 1:
                 audio = audio.mean(0).unsqueeze(0)
             audio = resample(audio, raw_sr, self.configs.sampling_rate, self.configs.device)
         else:
             audio = raw_audio.to(self.configs.device)
-            if audio.shape[0] == 2:
+            if audio.shape[0] > 1:
                 audio = audio.mean(0).unsqueeze(0)
 
         maxx = audio.abs().max()

GPT_SoVITS/TTS_infer_pack/TextPreprocessor.py

@@ -206,6 +206,7 @@ def clean_text_inf(self, text: str, language: str, version: str = "v2"):
         phones = cleaned_text_to_sequence(phones, version)
         return phones, word2ph, norm_text
 
+    @torch.no_grad()
     def get_bert_inf(self, phones: list, word2ph: list, norm_text: str, language: str):
         language = language.replace("all_", "")
         if language == "zh":