diff --git a/GPT_SoVITS/stream_v2pro.py b/GPT_SoVITS/stream_v2pro.py
index 718ef9af..7cd4351e 100644
--- a/GPT_SoVITS/stream_v2pro.py
+++ b/GPT_SoVITS/stream_v2pro.py
@@ -444,6 +444,7 @@ def export_prov2(
     output_path,
     device="cpu",
     is_half=True,
+    lang="auto",
 ):
     if export_torch_script.sv_cn_model == None:
         init_sv_cn(device,is_half)
@@ -454,7 +455,7 @@ def export_prov2(
     print(f"device: {device}")
 
     ref_seq_id, ref_bert_T, ref_norm_text = get_phones_and_bert(
-        ref_text, "all_zh", "v2"
+        ref_text, lang, "v2"
     )
     ref_seq = torch.LongTensor([ref_seq_id]).to(device)
     ref_bert = ref_bert_T.T
@@ -503,9 +504,9 @@ def export_prov2(
     stream_t2s = torch.jit.script(stream_t2s)
 
     ref_audio_sr = resamplex(ref_audio, 16000, 32000)
+    ref_audio_sr = ref_audio_sr.to(device)
     if is_half:
         ref_audio_sr = ref_audio_sr.half()
-    ref_audio_sr = ref_audio_sr.to(device)
 
     top_k = 15
 
@@ -588,6 +589,7 @@ if __name__ == "__main__":
     parser.add_argument("--device", help="Device to use", default="cuda" if torch.cuda.is_available() else "cpu")
     parser.add_argument("--version", help="version of the model", default="v2Pro")
     parser.add_argument("--no-half", action="store_true", help = "Do not use half precision for model weights")
+    parser.add_argument("--lang", default="auto", help="Language for text processing (default: auto)")
 
     args = parser.parse_args()
 
@@ -605,4 +607,5 @@ if __name__ == "__main__":
             output_path=args.output_path,
             device=args.device,
             is_half=is_half,
+            lang=args.lang,
         )