From e35591f11348a67ac8312990a6bb58e0a0602046 Mon Sep 17 00:00:00 2001
From: FelysNeko <88437453+FelysNeko@users.noreply.github.com>
Date: Thu, 9 Jul 2026 09:53:04 -0400
Subject: [PATCH] fix: Qwen3-TTS project input text embedding

---
 swift/model/models/qwen.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/swift/model/models/qwen.py b/swift/model/models/qwen.py
index 209d5bcdda..0e6ebea275 100644
--- a/swift/model/models/qwen.py
+++ b/swift/model/models/qwen.py
@@ -1793,7 +1793,8 @@ def tts_forward(self,
         input_codec_ids = input_ids[:, :, 1]
 
         # Build text and codec embeddings
-        input_text_embedding = self.talker.model.text_embedding(input_text_ids) * text_embedding_mask
+        input_text_embedding = self.talker.text_projection(
+            self.talker.model.text_embedding(input_text_ids)) * text_embedding_mask
         input_codec_embedding = self.talker.model.codec_embedding(input_codec_ids) * codec_embedding_mask
         # Inject speaker embedding at position 6
         input_codec_embedding[:, 6, :] = speaker_embedding