diff --git a/GPT_SoVITS/AR/models/t2s_model.py b/GPT_SoVITS/AR/models/t2s_model.py
index ac905f4b..7ed63e22 100644
--- a/GPT_SoVITS/AR/models/t2s_model.py
+++ b/GPT_SoVITS/AR/models/t2s_model.py
@@ -553,6 +553,10 @@ class Text2SemanticDecoder(nn.Module):
                 mask=xy_attn_mask,
             )
             logits = self.ar_predict_layer(xy_dec[:, -1])
+
+            eos_penalty = 2.0
+            logits[:, self.EOS] -= eos_penalty
+
             samples = topk_sampling(logits, top_k=top_k, top_p=1.0, temperature=temperature)
 
             if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
@@ -705,6 +709,9 @@ class Text2SemanticDecoder(nn.Module):
                 xy_dec, k_cache, v_cache = self.t2s_transformer.decode_next_token(xy_pos, k_cache, v_cache, attn_mask)
             logits = self.ar_predict_layer(xy_dec[:, -1])
 
+            eos_penalty = 2.0
+            logits[:, self.EOS] -= eos_penalty
+
             if idx == 0:
                 attn_mask = F.pad(attn_mask[:, :, -1].unsqueeze(-2), (0, 1), value=False)
             else:
@@ -895,6 +902,9 @@ class Text2SemanticDecoder(nn.Module):
 
             logits = self.ar_predict_layer(xy_dec[:, -1])
 
+            eos_penalty = 2.0
+            logits[:, self.EOS] -= eos_penalty
+
             if idx == 0:
                 xy_attn_mask = None
             if idx < 11:  ###至少预测出10个token不然不给停止（0.4s）