diff --git a/src/ComicInfoBuilder.py b/src/ComicInfoBuilder.py
index 70f6bd4..b93e969 100644
--- a/src/ComicInfoBuilder.py
+++ b/src/ComicInfoBuilder.py
@@ -439,8 +439,7 @@ class ComicInfoBuilder:
         # ----- Title / Series -----------------------------------------------
         add("Title", sd.get("Title") or f"Chapter {self._chapter}")
         add("Series", md.get("title") or self._manga_title)
-        add("LocalizedSeries",
-            md.get("native_title") or md.get("romanized_title"))
+        add("LocalizedSeries", self._romanized_for_native(md))
         add("SeriesSort", self._get_sort_title(md))
         add("Number", sd.get("Number") or self._chapter)
         add("Count", md.get("total_chapters"))
@@ -648,6 +647,70 @@ class ComicInfoBuilder:
     # ======================================================================
     # Title helpers
     # ======================================================================
+    # Mapping from series type to the matching romanized language code(s)
+    # in the MangaBaka titles array.  Used to pick the correct romaji /
+    # romaja / pinyin for LocalizedSeries.
+    _ROMANIZED_LANG_BY_TYPE = {
+        "manga":  ("ja-latn", "ja-romaji"),
+        "manhwa": ("ko-latn", "ko-romaji"),
+        "manhua": ("zh-latn",),
+    }
+
+    @classmethod
+    def _romanized_for_native(cls, md: dict) -> "str | None":
+        """
+        Picks the romanized title in the manga's original language from the
+        ``titles`` array.
+
+        The series' original language is inferred from ``type``::
+
+            manga  -> ja-Latn  (Japanese romaji)
+            manhwa -> ko-Latn  (Korean romaja)
+            manhua -> zh-Latn  (Chinese pinyin)
+
+        Among multiple entries for the matching language, the one with the
+        highest "quality score" wins (``official`` trait > ``is_primary`` >
+        first seen).
+
+        The root-level ``romanized_title`` field is **deliberately not used
+        as a fallback** — MangaBaka frequently stores a different language's
+        romanization there (e.g. Korean romaja on a Japanese manga), which
+        is exactly what this function is meant to avoid.
+
+        Returns ``None`` when no romanized title is available for the
+        inferred language.
+        """
+        mtype = (md.get("type") or "").lower()
+        langs = cls._ROMANIZED_LANG_BY_TYPE.get(mtype)
+        if not langs:
+            return None
+
+        titles = md.get("titles") or md.get("alt_titles") or []
+        if not isinstance(titles, list):
+            return None
+
+        best_score = -1
+        best_title: "str | None" = None
+        for entry in titles:
+            if not isinstance(entry, dict):
+                continue
+            lang = (entry.get("language") or entry.get("lang") or "").lower()
+            if lang not in langs:
+                continue
+            title = entry.get("title")
+            if not title:
+                continue
+            traits = entry.get("traits") or []
+            score = 0
+            if "official" in traits:
+                score += 2
+            if entry.get("is_primary"):
+                score += 1
+            if score > best_score:
+                best_score = score
+                best_title = title
+        return best_title
+
     def _get_sort_title(self, md: dict) -> "str | None":
         """
         Returns the SeriesSort title in the configured language.