Normalize question marks before local TTS

platform/docker/tts-gateway/test_tts_gateway.py

@@ -120,6 +120,11 @@ class LanguageSegmentationTests(unittest.TestCase):
         )
         self.assertNotRegex(spoken, r"[:;\-‐‑‒–—−]")
 
+    def test_question_and_exclamation_marks_are_not_sent_to_xtts(self):
+        spoken = gateway.clean_for_speech("Wie geht es dir? Wirklich gut!")
+        self.assertEqual(spoken, "Wie geht es dir. Wirklich gut.")
+        self.assertNotRegex(spoken, r"[!?]")
+
     def test_punctuation_only_segments_are_never_synthesized(self):
         segments = gateway.prepare_segments("Status: – alles läuft.")
         self.assertTrue(segments)

platform/docker/tts-gateway/tts_gateway.py

@@ -298,6 +298,10 @@ def clean_for_speech(text: str) -> str:
     text = re.sub(r"(?<=\w)[\-‐‑‒–—−](?=\w)", " ", text)
     text = re.sub(r"\s+[\-‐‑‒–—−]\s+", ", ", text)
     text = re.sub(r"\s*[:;]+\s*", ", ", text)
+    # XTTS can pronounce literal question/exclamation glyphs as short
+    # nonsense syllables (for example "?" as "nau"). Retain a sentence
+    # boundary for pacing, but never pass those glyphs to the model.
+    text = re.sub(r"\s*[!?]+\s*", ". ", text)
     text = "".join(
         char for char in text
         if unicodedata.category(char) not in {"So", "Cs"}