Reduce XTTS tail hallucinations

platform/docker/tts-gateway/test_tts_gateway.py

@@ -130,7 +130,7 @@ class LanguageSegmentationTests(unittest.TestCase):
         self.assertTrue(segments)
         self.assertTrue(all(any(char.isalnum() for char in part) for _, part in segments))
 
-    def test_weather_summary_is_split_into_complete_sentences(self):
+    def test_weather_summary_keeps_complete_sentences_together(self):
         text = (
             "Heute in Rastatt: teils sonnig, trocken, 10 bis 22 Grad. "
             "Abends wolkiger, 16 bis 21 Grad. Böen bis 29 km/h. "
@@ -138,13 +138,23 @@ class LanguageSegmentationTests(unittest.TestCase):
         )
         segments = gateway.prepare_segments(text)
         spoken = " ".join(part for _, part in segments)
-        self.assertEqual(len(segments), 5)
+        self.assertEqual(len(segments), 1)
         self.assertTrue(all(len(part) <= gateway.CHUNK_CHARS
                             for _, part in segments))
         self.assertIn("29 Kilometer pro Stunde", spoken)
         self.assertIn("Kein Regen erwartet", spoken)
         self.assertIn("wetteronline Punkt de", spoken)
 
+    def test_short_followup_sentence_shares_the_same_xtts_request(self):
+        text = (
+            "Ehrlich gesagt habe ich keine echten Gefühle wie Menschen, aber "
+            "ich bin wach, aufmerksam und motiviert, dir zu helfen. "
+            "Klingt das gut?"
+        )
+        segments = gateway.prepare_segments(text)
+        self.assertEqual(len(segments), 1)
+        self.assertIn("dir zu helfen. Klingt das gut.", segments[0][1])
+
     def test_normal_sentence_is_not_broken_into_word_sized_requests(self):
         text = (
             "Die automatische Komprimierung ist in OpenClaw standardmäßig "

platform/docker/tts-gateway/tts_gateway.py

@@ -318,12 +318,12 @@ def clean_for_speech(text: str) -> str:
 
 
 def _split_chunk(text: str, limit: int = CHUNK_CHARS) -> list[str]:
-    """Return sentence-sized XTTS requests with a conservative hard ceiling.
+    """Return paragraph-sized XTTS requests with a conservative hard ceiling.
 
-    A sentence is deliberately never combined with the following sentence.
-    Overlong sentences are split at clause boundaries first and at words only
-    as a last resort. This preserves XTTS prosody without exposing it to an
-    unbounded paragraph.
+    Complete neighbouring sentences are combined while they fit. This avoids
+    restarting the generative XTTS decoder after every short sentence, which
+    can create invented tail syllables between sentences. Overlong sentences
+    are split at clause boundaries first and at words only as a last resort.
     """
     text = text.strip()
     if not text:
@@ -331,38 +331,50 @@ def _split_chunk(text: str, limit: int = CHUNK_CHARS) -> list[str]:
 
     sentences = re.split(r"(?<=[.!?])\s+|\s+(?=\d+[.)]\s)", text)
     chunks: list[str] = []
+    current = ""
     for sentence in sentences:
         sentence = sentence.strip()
         if not sentence:
             continue
         if len(sentence) <= limit:
-            chunks.append(sentence)
+            candidate = f"{current} {sentence}".strip()
+            if current and len(candidate) > limit:
+                chunks.append(current)
+                current = sentence
+            else:
+                current = candidate
             continue
 
+        if current:
+            chunks.append(current)
+            current = ""
+
         # Retain commas in the preceding clause so XTTS can reproduce the
         # intended pause. Semicolons and colons were normalized earlier.
         clauses = re.split(r"(?<=,)\s+", sentence)
-        current = ""
+        long_current = ""
         for clause in clauses:
             clause = clause.strip()
-            candidate = f"{current} {clause}".strip()
-            if current and len(candidate) > limit:
-                chunks.append(current)
-                current = ""
+            candidate = f"{long_current} {clause}".strip()
+            if long_current and len(candidate) > limit:
+                chunks.append(long_current)
+                long_current = ""
             if len(clause) <= limit:
-                current = f"{current} {clause}".strip()
+                long_current = f"{long_current} {clause}".strip()
                 continue
 
             # A clause without a usable pause can still exceed the ceiling.
             for word in clause.split():
-                candidate = f"{current} {word}".strip()
-                if current and len(candidate) > limit:
-                    chunks.append(current)
-                    current = word
+                candidate = f"{long_current} {word}".strip()
+                if long_current and len(candidate) > limit:
+                    chunks.append(long_current)
+                    long_current = word
                 else:
-                    current = candidate
-        if current:
-            chunks.append(current)
+                    long_current = candidate
+        if long_current:
+            chunks.append(long_current)
+    if current:
+        chunks.append(current)
     return chunks