From 744a207e5a4c1dffd43c2646738c5df82e887922 Mon Sep 17 00:00:00 2001
From: Mikei386 <44135113+Mikei386@users.noreply.github.com>
Date: Fri, 4 Sep 2026 09:30:04 +0200
Subject: [PATCH] Stabilize XTTS sentence endings

---
 platform/docker/tts-gateway/test_tts_gateway.py | 10 ++++++++++
 platform/docker/tts-gateway/tts_gateway.py      | 15 ++++++++++++++-
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/platform/docker/tts-gateway/test_tts_gateway.py b/platform/docker/tts-gateway/test_tts_gateway.py
index eb3eba0..8d6993e 100644
--- a/platform/docker/tts-gateway/test_tts_gateway.py
+++ b/platform/docker/tts-gateway/test_tts_gateway.py
@@ -155,6 +155,16 @@ class LanguageSegmentationTests(unittest.TestCase):
         self.assertEqual(len(segments), 1)
         self.assertIn("dir zu helfen. Klingt das gut.", segments[0][1])
 
+    def test_xtts_receives_a_stable_terminal_stop_cue(self):
+        self.assertEqual(
+            gateway._stabilize_xtts_ending("Heute bleibt es trocken."),
+            "Heute bleibt es trocken;",
+        )
+        self.assertEqual(
+            gateway._stabilize_xtts_ending("Klingt das gut."),
+            "Klingt das gut;",
+        )
+
     def test_normal_sentence_is_not_broken_into_word_sized_requests(self):
         text = (
             "Die automatische Komprimierung ist in OpenClaw standardmäßig "
diff --git a/platform/docker/tts-gateway/tts_gateway.py b/platform/docker/tts-gateway/tts_gateway.py
index 4ad97a9..ef04587 100644
--- a/platform/docker/tts-gateway/tts_gateway.py
+++ b/platform/docker/tts-gateway/tts_gateway.py
@@ -419,10 +419,23 @@ def _speaker_conditioning() -> dict:
         return SPEAKER_CONDITIONING
 
 
+def _stabilize_xtts_ending(text: str) -> str:
+    """Give XTTS a reliable stop cue without changing the visible answer.
+
+    XTTS v2 can continue with invented syllables after a terminal full stop,
+    especially in German. A terminal semicolon is tokenized as a stronger,
+    more reliable boundary while retaining neutral sentence intonation.
+    """
+    spoken = text.rstrip()
+    if spoken.endswith((".", "!", "?", ";", ":")):
+        spoken = spoken[:-1].rstrip()
+    return f"{spoken};"
+
+
 def _xtts_pcm(text: str, language: str, conditioning: dict) -> bytes:
     payload = {
         **conditioning,
-        "text": text,
+        "text": _stabilize_xtts_ending(text),
         "language": language,
         "add_wav_header": True,
         "stream_chunk_size": "20",