diff --git a/.jules/bolt.md b/.jules/bolt.md
index 8513f7ad5..0a1774476 100644
--- a/.jules/bolt.md
+++ b/.jules/bolt.md
@@ -27,3 +27,7 @@
 ## 2024-05-18 - Optimize Regex pattern detection in DictionaryFacilitatorImpl
 **Learning:** `DictionaryFacilitatorImpl.isRegexPattern` used chained `String.contains()` and `String.startsWith()` calls, which internally execute full loops over the string (or delegate to regex). Replacing these chained calls with a single manual O(N) iteration reduces overhead and prevents repeatedly traversing the string.
 **Action:** Replace chained string detection methods (`contains`, `startsWith`, `endsWith`) with a single manual loop over the string's indices when checking multiple characters in performance hot paths.
+
+## 2024-05-18 - Optimize Regex split overhead
+**Learning:** `LanguageDetector.detectByHeuristics` used `text.lowercase().split(Regex("[^\\p{L}]+")).filter { it.length >= 2 }`. While the Regex was pre-compiled, the split and filter operations still allocated intermediate lists and required regex engine execution, causing GC pressure in a frequently invoked text processing path.
+**Action:** Replace `Regex.split` combined with `.filter` in hot paths with manual O(N) string index iteration using character checks (like `char.isLetter()`) and a `StringBuilder` to eliminate regex execution and avoid allocating intermediate collections.
diff --git a/app/src/main/java/helium314/keyboard/latin/utils/LanguageDetector.kt b/app/src/main/java/helium314/keyboard/latin/utils/LanguageDetector.kt
index 2be7e5f72..1419fcaa0 100644
--- a/app/src/main/java/helium314/keyboard/latin/utils/LanguageDetector.kt
+++ b/app/src/main/java/helium314/keyboard/latin/utils/LanguageDetector.kt
@@ -148,8 +148,6 @@ object LanguageDetector {
         return null
     }
 
-    private val NON_LETTER_REGEX = Regex("[^\\p{L}]+")
-
     private fun detectByHeuristics(text: String, targetLangCode: String?): String? {
         val scores = mutableMapOf<String, Int>()
 
@@ -175,18 +173,35 @@ object LanguageDetector {
             }
         }
 
-        val words = text.lowercase().split(NON_LETTER_REGEX).filter { it.length >= 2 }
-        for (word in words) {
-            if (SPANISH_WORDS.contains(word)) addScore("es", 2)
-            if (FRENCH_WORDS.contains(word)) addScore("fr", 2)
-            if (GERMAN_WORDS.contains(word)) addScore("de", 2)
-            if (ITALIAN_WORDS.contains(word)) addScore("it", 2)
-            if (PORTUGUESE_WORDS.contains(word)) addScore("pt", 2)
-            if (DUTCH_WORDS.contains(word)) addScore("nl", 2)
-            if (TURKISH_WORDS.contains(word)) addScore("tr", 2)
-            if (POLISH_WORDS.contains(word)) addScore("pl", 2)
-            if (ENGLISH_WORDS.contains(word)) addScore("en", 2)
+        val textLower = text.lowercase()
+        val len = textLower.length
+        val wordBuilder = StringBuilder()
+
+        fun processWord() {
+            if (wordBuilder.length >= 2) {
+                val word = wordBuilder.toString()
+                if (SPANISH_WORDS.contains(word)) addScore("es", 2)
+                if (FRENCH_WORDS.contains(word)) addScore("fr", 2)
+                if (GERMAN_WORDS.contains(word)) addScore("de", 2)
+                if (ITALIAN_WORDS.contains(word)) addScore("it", 2)
+                if (PORTUGUESE_WORDS.contains(word)) addScore("pt", 2)
+                if (DUTCH_WORDS.contains(word)) addScore("nl", 2)
+                if (TURKISH_WORDS.contains(word)) addScore("tr", 2)
+                if (POLISH_WORDS.contains(word)) addScore("pl", 2)
+                if (ENGLISH_WORDS.contains(word)) addScore("en", 2)
+            }
+            wordBuilder.clear()
+        }
+
+        for (i in 0 until len) {
+            val char = textLower[i]
+            if (char.isLetter()) {
+                wordBuilder.append(char)
+            } else {
+                processWord()
+            }
         }
+        processWord()
 
         if (scores.isEmpty()) return null