Merge branch 'dev' into main

Author: wannaphong

CHANGELOG.md

@@ -17,6 +17,12 @@ and this project adheres to
 - Full release notes: <https://github.com/PyThaiNLP/pythainlp/releases>
 - Commit history: <https://github.com/PyThaiNLP/pythainlp/compare/v5.3.3...v5.3.4>
 
+## [Unreleased]
+
+## Changed
+
+- Improve guardrails in `check_sara()` and `nighit()`
+
 ## [5.3.5] - 2026-07-30
 
 ### Fixed
@@ -26,7 +32,6 @@ and this project adheres to
 - Fix: CodeQL ReDoS vulnerability in ULMFiT replace_url #1400
 - fix: remove unconditional overwrite that disables teacher forcing in Seq2Seq #1380
 
-
 ## [5.3.4] - 2026-04-02
 
 ### Fixed
@@ -36,6 +41,18 @@ and this project adheres to
 - Build WSD Trie after populating dictionary (#1388).
 - Doctests across all modules (#1392).
 
+## [Unreleased]
+
+### Changed
+
+- `pythainlp.tokenize.deepcut`: migrated from the TensorFlow-based `deepcut`
+  package to a built-in ONNX inference engine, removing the TensorFlow
+  dependency. The `deepcut.onnx` model (ported from
+  [LEKCut](https://github.com/PyThaiNLP/LEKCut)) is now bundled with PyThaiNLP.
+  The `segment()` API is unchanged; the `custom_dict` parameter is kept for
+  backward compatibility but is no longer applied to the model inference.
+  Deepcut tests moved from `tests/noauto_tensorflow/` to `tests/noauto_onnx/`.
+
 ## [5.3.3] - 2026-03-26
 
 Security fixes and thai2rom_onnx bug fixes.

docs/api/transliterate.rst

@@ -17,16 +17,25 @@ Modules
 
     The `transliterate` function serves as a versatile transliteration tool, offering a range of transliteration engines to choose from. It provides flexibility and customization for your transliteration needs.
 
+.. autofunction:: transliterate_wiktionary
+    :noindex:
+
 .. autofunction:: pronunciate
     :noindex:
 
     This function provides assistance in generating phonetic representations of Thai words, which is particularly useful for language learning and pronunciation practice.
 
+.. autofunction:: pronunciate_pali
+    :noindex:
+
 .. autofunction:: puan
     :noindex:
 
     The `puan` function offers a unique transliteration feature known as "Puan." It provides a specialized transliteration method for Thai text and is an additional option for rendering Thai text into English characters.
 
+.. autofunction:: get_word_dict
+    :noindex:
+
 .. autoclass:: pythainlp.transliterate.wunsen.WunsenTransliterate
    :members:
 

pyproject.toml

@@ -219,7 +219,6 @@ noauto-torch = [
 
 # TensorFlow-based dependencies - for tests.noauto_tensorflow
 noauto-tensorflow = [
-    "deepcut>=0.7.0",
     "numpy>=1.26.0",
 ]
 
@@ -247,10 +246,8 @@ full = [
     "attacut==1.0.6",
     "attaparse==1.0.0",
     "bpemb>=0.3.6,<0.4",
-    "budoux==0.8.4",
-    "deepcut==0.7.0.0",
-    "emoji>=0.6.0,<1",
-    "epitran==1.35.2",
+    "budoux==0.7.0",
+    "epitran==1.26.0",
     "esupar>=1.3.9,<2",
     'fairseq>=0.10.0,<0.13;python_version<"3.11"',
     'fairseq-fixed==0.12.3.1,<0.13;python_version>="3.11"',
@@ -455,7 +452,6 @@ module = [
     "attaparse.*",
     "bpemb.*",
     "budoux.*",
-    "deepcut.*",
     "emoji.*",
     "epitran.*",
     "esupar.*",

pythainlp/corpus/deepcut.onnx

@@ -1,4 +1,29 @@
 {
+    "deepcut_onnx": {
+        "name": "deepcut_onnx",
+        "latest_version": "1.0.0",
+        "description": "DeepCut ONNX model",
+        "long_description": "DeepCut Thai word segmentation model in ONNX format, ported from the original TensorFlow model",
+        "url": "https://github.com/PyThaiNLP/LEKCut",
+        "authors": [
+            "Rakpong Kittinaradorn",
+            "Titipat Achakulvisut",
+            "Korakot Chaovavanich",
+            "Kittinan Srithaworn",
+            "Pattarawat Chormai",
+            "Chanwit Kaewkasi",
+            "Tulakan Ruangrong",
+            "Krichkorn Oparad"
+        ],
+        "license": "MIT",
+        "versions": {
+            "1.0.0": {
+                "filename": "deepcut.onnx",
+                "md5": "f4662560dd9a706bfb1d7790ad6c667f",
+                "pythainlp_version": ">=5.4.0"
+            }
+        }
+    },
     "thainer": {
         "name": "thainer",
         "latest_version": "1.5.1",

pythainlp/corpus/default_db.json

@@ -11552,6 +11552,7 @@
 เค
 แคเฟ
 แคแกล
+แคคตัส
 แคชเชียร์
 แคชเมียร์
 แคช
@@ -18381,6 +18382,7 @@
 โดกเดก
 โด่งดัง
 โด่ง
+โดจิน
 โดซีแท็กเซล
 โดด ๆ
 โดดเด่น
@@ -20307,6 +20309,7 @@
 ตาลุก
 ตาลุชะ
 ตาลุ
+ตาลโตนด
 ตาเล็กตาน้อย
 ต๋าว
 ตาว
@@ -38600,6 +38603,7 @@
 มั่ง
 มั้ง
 มัง
+มังงะ
 มั่งมีในใจ แล่นใบบนบก
 มั่งมี
 มังราย
@@ -45834,6 +45838,7 @@
 ลิงหลอกเจ้า
 ลิงอุรังอุตัง
 ลิจฉวี
+ลิซ่า
 ลิดตีนปู
 ลิดรอน
 ลิด

pythainlp/corpus/words_th.txt

@@ -53,9 +53,15 @@ def check_sara(self, word: str) -> str:
         sara = []
         countoa = 0
 
+        if not word:
+            return ""
+
         # In case of การันย์
         if "์" in word[-1]:
             word = word[:-2]
+            # After removing the karun, the word may become empty (e.g. "ก์")
+            if not word:
+                return ""
 
         # In case of สระเดี่ยว
         for i in word:
@@ -217,7 +223,7 @@ def check_sara(self, word: str) -> str:
             sara.append("เอือ")
 
         if not sara:
-            return "Can't find Sara in this word"
+            return ""
 
         return sara[0]
 
@@ -251,6 +257,9 @@ def check_marttra(self, word: str) -> str:
         word = self.handle_karun_sound_silence(word)
         word = remove_tonemark(word)
 
+        if not word:
+            return ""
+
         # Check for ำ at the end (represents "am" sound, ends with m)
         if word[-1] == "ำ":
             return "กม"

pythainlp/khavee/core.py

@@ -35,14 +35,28 @@ def nighit(w1: str, w2: str) -> str:
         >>> nighit("สํ", "โยค")
         'สังโยค'
     """
+    if not isinstance(w1, str) or not isinstance(w2, str):
+        raise TypeError("Both w1 and w2 must be strings.")
+    w1 = w1.strip()
+    w2 = w2.strip()
+    if not w1:
+        return w2
+    if not w2:
+        return w1
     if not str(w1).endswith("ํ") and len(w1) != 2:
         raise NotImplementedError(f"The function doesn't support {w1}.")
     list_w1 = list(w1)
     list_w2 = list(w2)
     newword = []
     newword.append(list_w1[0])
     newword.append("ั")
-    consonant_start = [i for i in list_w2 if i in set(thai_consonants)][0]
+    _consonants = set(thai_consonants)
+    consonants_in_w2 = [i for i in list_w2 if i in _consonants]
+    if not consonants_in_w2:
+        raise ValueError(
+            f"w2 {w2!r} contains no Thai consonants."
+        )
+    consonant_start = consonants_in_w2[0]
     if consonant_start in ["ก", "ช", "ค", "ข", "ง"]:
         newword.append("ง")
     elif consonant_start in ["จ", "ฉ", "ช", "ฌ"]:

pythainlp/morpheme/word_formation.py

@@ -16,7 +16,7 @@
 _CONS_THANTHAKHAT: str = "กขฃคฅฆงจฉชซฌญฎฏฐฑฒณดตถทธนบปผฝพฟภมยรลวศษสหฬอฮ์"
 _THANTHAKHAT: str = "์"  # \u0e4c
 _C1: str = "กขฃคฆฅ"  # sound K -> coded letter 1
-_C2: str = "จฉชฌซฐทฒดฎตสศษ"  # D -> 2
+_C2: str = "จฉชฌซฐฏทฑฒถธดฎตสศษ"  # D -> 2
 _C3: str = "ฟฝพผภบป"  # B -> 3
 _C4: str = "ง"  # NG -> 4
 _C5: str = "ลฬรนณฦญ"  # N -> 5
@@ -71,7 +71,8 @@ def metasound(text: str, length: int = 4) -> str:
             chars[i] = " "
         i += 1
 
-    # retain first consonant, encode the rest
+    # filter out spaces left by karan removal, then truncate
+    chars = [c for c in chars if c != " "]
     chars = chars[:length]
     i = 1
     while i < len(chars):

pythainlp/soundex/metasound.py

@@ -18,7 +18,7 @@
 
 _C0: str = "AEIOUHWYอ"
 _C1: str = "BFPVบฝฟปผพภว"
-_C2: str = "CGJKQSXZขฃคฅฆฉขฌกจซศษส"
+_C2: str = "CGJKQSXZขฃคฅฆฉฌกจซศษส"
 _C3: str = "DTฎดฏตฐฑฒถทธ"
 _C4: str = "Lลฬ"
 _C5: str = "MNมณน"