From d08214dd22e8837b4f9eb92f94a1d419806e182a Mon Sep 17 00:00:00 2001
From: ChasonJiang <1440499136@qq.com>
Date: Tue, 1 Jul 2025 22:27:03 +0800
Subject: [PATCH] 	modified:   GPT_SoVITS/TTS_infer_pack/TTS.py

---
 GPT_SoVITS/TTS_infer_pack/TTS.py | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/GPT_SoVITS/TTS_infer_pack/TTS.py b/GPT_SoVITS/TTS_infer_pack/TTS.py
index d2a2d3ce..813117a2 100644
--- a/GPT_SoVITS/TTS_infer_pack/TTS.py
+++ b/GPT_SoVITS/TTS_infer_pack/TTS.py
@@ -1799,9 +1799,16 @@ class TTS:
             f2 = audio_fragments[i + 1]
             w1 = f1[-overlap_len:]
             w2 = f2[:overlap_len]
-            assert w1.shape == w2.shape
-            corr = F.conv1d(w1.view(1, 1, -1), w2.view(1, 1, -1), padding=w2.shape[-1] // 2).view(-1)[:-1]
-            idx = corr.argmax()
+            w2 = w2[-w2.shape[-1]//2:]
+            # assert w1.shape == w2.shape
+            corr = F.conv1d(w1.view(1, 1, -1), w2.view(1, 1, -1)).view(-1)
+
+            squared_sum = F.conv1d(w1.view(1, 1, -1)**2, torch.ones_like(w2).view(1, 1, -1)).view(-1)+ 1e-8
+            idx = (corr/squared_sum.sqrt()).argmax()
+
+            print(f"seg_idx: {idx}")
+
+            # idx = corr.argmax()
             f1_ = f1[: -(overlap_len - idx)]
             audio_fragments[i] = f1_