diff --git a/style_bert_vits2/models/infer.py b/style_bert_vits2/models/infer.py
index 835a042..c07b8c0 100644
--- a/style_bert_vits2/models/infer.py
+++ b/style_bert_vits2/models/infer.py
@@ -224,14 +224,23 @@ def infer(
         en_bert = en_bert[:, :-2]
 
     with torch.no_grad():
+        # BERT may run in fp16 (for example on Colab) while a loaded TTS model
+        # remains fp32. Conv1d/Linear require inputs and parameters to use the
+        # same dtype outside autocast, so normalize all floating conditioning
+        # tensors to the TTS model dtype at this boundary.
+        model_dtype = next(net_g.parameters()).dtype
         x_tst = phones.to(device).unsqueeze(0)
         tones = tones.to(device).unsqueeze(0)
         lang_ids = lang_ids.to(device).unsqueeze(0)
-        bert = bert.to(device).unsqueeze(0)
-        ja_bert = ja_bert.to(device).unsqueeze(0)
-        en_bert = en_bert.to(device).unsqueeze(0)
+        bert = bert.to(device=device, dtype=model_dtype).unsqueeze(0)
+        ja_bert = ja_bert.to(device=device, dtype=model_dtype).unsqueeze(0)
+        en_bert = en_bert.to(device=device, dtype=model_dtype).unsqueeze(0)
         x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
-        style_vec_tensor = torch.from_numpy(style_vec).to(device).unsqueeze(0)
+        style_vec_tensor = (
+            torch.from_numpy(style_vec)
+            .to(device=device, dtype=model_dtype)
+            .unsqueeze(0)
+        )
         del phones
         sid_tensor = torch.LongTensor([sid]).to(device)