Speak aspect ratios correctly in TTS

platform/docker/tts-gateway/test_tts_gateway.py

@@ -132,6 +132,25 @@ class LanguageSegmentationTests(unittest.TestCase):
         self.assertIn("5 bis 6 Uhr", spoken)
         self.assertIn("Wind maximal 16 Kilometer pro Stunde", spoken)
 
+    def test_qwen_speaks_aspect_ratios_as_ratios(self):
+        spoken = gateway.prepare_for_qwen_speech(
+            "Cover sind im Hochformat (2:3), Screenshots im Querformat "
+            "(16:9), ein Quadrat im Seitenverhältnis 2:2 und 4:3-Format."
+        )
+        self.assertIn("Hochformat (2 zu 3)", spoken)
+        self.assertIn("Querformat (16 zu 9)", spoken)
+        self.assertIn("Seitenverhältnis 2 zu 2", spoken)
+        self.assertIn("4 zu 3-Format", spoken)
+
+    def test_qwen_keeps_clock_times_distinct_from_aspect_ratios(self):
+        spoken = gateway.prepare_for_qwen_speech(
+            "Beginn um 16:09 Uhr, Fehler um 02:14; das Videoformat ist 16:9."
+        )
+        self.assertIn("16 Uhr 9", spoken)
+        self.assertNotIn("16 Uhr 9 Uhr", spoken)
+        self.assertIn("2 Uhr 14", spoken)
+        self.assertIn("Videoformat ist 16 zu 9", spoken)
+
     def test_qwen_speaks_strict_date_ranges_as_calendar_dates(self):
         spoken = gateway.prepare_for_qwen_speech(
             "Neuigkeiten vom 04.–05.09. und Vergleich 04.09.–06.10.2026."

platform/docker/tts-gateway/tts_gateway.py

@@ -269,6 +269,36 @@ def _spoken_ipv4(match: re.Match) -> str:
     return " Punkt ".join(str(int(part)) for part in match.group(0).split("."))
 
 
+def _normalize_aspect_ratios(text: str) -> str:
+    """Speak colon notation as a ratio only when the surrounding text says so.
+
+    A bare ``16:09`` remains a clock time.  This deliberately avoids a global
+    replacement of common ratios because ``16:9`` can also be a valid time.
+    """
+    cue = (
+        r"(?:Seitenverh[aä]ltnis|Bildseitenverh[aä]ltnis|Bildformat|"
+        r"Videoformat|Hochformat|Querformat|Format|Aspect[- ]?Ratio)"
+    )
+    text = re.sub(
+        rf"\b({cue}\b(?:\s+(?:von|im|ist|betr[aä]gt))?\s*[\(\[]?\s*)"
+        rf"(\d{{1,3}})\s*:\s*(\d{{1,3}})",
+        lambda match: (
+            f"{match.group(1)}{int(match.group(2))} zu {int(match.group(3))}"
+        ),
+        text,
+        flags=re.IGNORECASE,
+    )
+    return re.sub(
+        rf"\b(\d{{1,3}})\s*:\s*(\d{{1,3}})"
+        rf"(\s*[-‐‑‒–—−]?\s*{cue}\b)",
+        lambda match: (
+            f"{int(match.group(1))} zu {int(match.group(2))}{match.group(3)}"
+        ),
+        text,
+        flags=re.IGNORECASE,
+    )
+
+
 def normalize_for_german_speech(text: str) -> str:
     """Turn common visual notation into unambiguous spoken German."""
     text = re.sub(r"\bv\.\s*a\.", "vor allem", text, flags=re.IGNORECASE)
@@ -281,10 +311,14 @@ def normalize_for_german_speech(text: str) -> str:
         _spoken_ipv4,
         text,
     )
+    # A colon is ambiguous between an aspect ratio and a clock time. Resolve
+    # ratios first, but only when an explicit format cue is present.
+    text = _normalize_aspect_ratios(text)
     text = re.sub(
-        r"\b([01]?\d|2[0-3]):([0-5]\d)\b",
+        r"\b([01]?\d|2[0-3]):([0-5]\d)\b(?:\s*Uhr\b)?",
         lambda match: f"{int(match.group(1))} Uhr {int(match.group(2))}",
         text,
+        flags=re.IGNORECASE,
     )
     text = re.sub(
         r"\b([01]?\d|2[0-3])\s*[-‐‑‒–—−]\s*"