Update imports to use `GPT_SoVITS` module paths and replace `torchaudio` with `librosa` for loading reference audio

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -21,20 +21,20 @@ import numpy as np
 import torch
 import torch.nn.functional as F
 import yaml
-from AR.models.t2s_lightning_module import Text2SemanticLightningModule
-from BigVGAN.bigvgan import BigVGAN
-from feature_extractor.cnhubert import CNHubert
-from module.mel_processing import mel_spectrogram_torch, spectrogram_torch
-from module.models import SynthesizerTrn, SynthesizerTrnV3, Generator
+from GPT_SoVITS.AR.models.t2s_lightning_module import Text2SemanticLightningModule
+from GPT_SoVITS.BigVGAN.bigvgan import BigVGAN
+from GPT_SoVITS.feature_extractor.cnhubert import CNHubert
+from GPT_SoVITS.module.mel_processing import mel_spectrogram_torch, spectrogram_torch
+from GPT_SoVITS.module.models import SynthesizerTrn, SynthesizerTrnV3, Generator
 from peft import LoraConfig, get_peft_model
-from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
+from GPT_SoVITS.process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
 from transformers import AutoModelForMaskedLM, AutoTokenizer
 
 from tools.audio_sr import AP_BWE
 from tools.i18n.i18n import I18nAuto, scan_language_list
-from TTS_infer_pack.text_segmentation_method import splits
-from TTS_infer_pack.TextPreprocessor import TextPreprocessor
-from sv import SV
+from GPT_SoVITS.TTS_infer_pack.text_segmentation_method import splits
+from GPT_SoVITS.TTS_infer_pack.TextPreprocessor import TextPreprocessor
+from GPT_SoVITS.sv import SV
 
 resample_transform_dict = {}
 
@@ -770,7 +770,12 @@ class TTS:
             self.prompt_cache["refer_spec"][0] = spec_audio
 
     def _get_ref_spec(self, ref_audio_path):
-        raw_audio, raw_sr = torchaudio.load(ref_audio_path)
+        raw_audio_np, raw_sr = librosa.load(ref_audio_path, sr=None, mono=False)
+        # Convert numpy array to torch tensor and ensure shape matches torchaudio output
+        if raw_audio_np.ndim == 1:
+            raw_audio = torch.from_numpy(raw_audio_np).unsqueeze(0)
+        else:
+            raw_audio = torch.from_numpy(raw_audio_np)
         raw_audio = raw_audio.to(self.configs.device).float()
         self.prompt_cache["raw_audio"] = raw_audio
         self.prompt_cache["raw_sr"] = raw_sr