phonsobon/khmer-G2P

Khmer G2P — word → pronunciation

A clean, ready-to-train dataset of 38,629 Khmer word → pronunciation pairs (37,755 unique words) for grapheme-to-phoneme (G2P) models: given how a Khmer word is written, predict how it is pronounced.

The pronunciations come from the Khmer Dictionary 2022 (វចនានុក្រមខ្មែរ ២០២២) of the National Council of Khmer Language (NCKL), Royal Academy of Cambodia. The dictionary writes each word's pronunciation in Khmer script, syllable by syllable, which shows how the word is actually read:

word (written)	pronunciation (spoken)	what it shows
ក្របី	ក្រ-បី	syllable boundaries
កម្ពុជា	កាំ-ពុ-ជា	ម្ព is read as nasal ំ + ព
ប្រទេស	ប្រ-ទេស
វិទ្យាល័យ	វិត-ទ្យា-ឡៃ	ទ closes the first syllable and starts the next; ល័យ → ឡៃ
កាលវិបត្តិ	កា-ឡៈ-វិ-ប័ត or កាល-វិ-ប័ត	two accepted readings
សាស្ត្រ	សាស	silent final letters
ការណ៍	កា	silent letters marked with ៍

This is not the original dictionary. It is derived from it and cleaned for model training: entries were removed, readings were split and source typos were fixed, as documented below. For the full dictionary with definitions, see phonsobon/khmer-khmer-dictionary.

Quick start

from datasets import load_dataset

ds = load_dataset("phonsobon/khmer-G2P")
print(ds)
# DatasetDict({
#     train:      Dataset({features: ['word', 'pronunciation', 'syllables'], num_rows: 34770})
#     validation: Dataset({features: ['word', 'pronunciation', 'syllables'], num_rows: 1939})
#     test:       Dataset({features: ['word', 'pronunciation', 'syllables'], num_rows: 1920})
# })
ds["train"][0]
# {'word': 'កកិត', 'pronunciation': 'ក-កិត', 'syllables': 2}

A complete training example (a 7.5M-parameter character-level Transformer, with a Google Colab notebook) is at github.com/PhonSobon/Khmer-pronunciation.

Data fields

field	type	description
word	string	The Khmer word as written. Khmer script only (U+1780–U+17DD): no spaces, digits or punctuation.
pronunciation	string	NCKL Khmer-script respelling, syllables separated by -. Khmer script and - only.
syllables	int	Number of syllables (pronunciation.count("-") + 1).

Character sets. Words use 75 distinct characters, pronunciations 64. The only character in pronunciations that never appears in words is the syllable separator -. Pronunciations also use the diacritics ៉ / ៊ (register shift) and ៈ (short final vowel) much more often than spelling does, e.g. ឧរគ → អ៊ុ-រ៉ៈ-គៈ.

Splits

split	pairs	unique words
train	34,770	33,979
validation	1,939	1,888
test	1,920	1,888
total	38,629	37,755

Words are shuffled with a fixed seed (42) and split 90 / 5 / 5 by word, not by row. All readings of a word are always in the same split, so no test word is ever seen in training.

Statistics

Syllables per pronunciation

syllables	1	2	3	4	5	6	7	8	9+
pairs	4,691	12,194	8,755	7,297	3,679	1,381	424	149	59

• Average word length: 7.8 characters (longest: 39). Average pronunciation length: 10.3 characters (longest: 50).
• Longest entry: ម៉ាស៊ីនប្រតិបត្តិការធនាគារស្វ័យប្រវត្ត → ម៉ា-ស៊ីន-ប្រៈ-តិ-ប័ត-តិ-កា-ធៈ-នា-គា-ស្វៃ-ប្រ-វ័ត (13 syllables).
• 3,879 pairs are read exactly as written (one-syllable words such as ប្រុស → ប្រុស).

Multiple readings. 869 words have more than one accepted pronunciation (one row per reading, at most 3), e.g. អន្តរ → អ័ន-ដ | អន់-តៈ-រ៉ៈ | អន់-ដ. When evaluating, count a prediction as correct if it matches any reference reading of the word:

from collections import defaultdict
refs = defaultdict(set)
for r in ds["test"]:
    refs[r["word"]].add(r["pronunciation"])
# word accuracy = mean(pred[w] in refs[w] for w in refs)

How the dataset was built

Source: the raw NCKL Khmer Dictionary 2022 (44,698 rows). Every step is in scripts/build_g2p_dataset.py.

step	pairs	change
Word / pronunciation pairs from the dictionary (each sense row inherits its entry's word and pronunciation; sub-entries use their own word)	38,725
Removed: no pronunciation (idioms, proverbs, sayings)	38,562	−163
Alternative readings [a] ឬ [b] split into separate rows	39,418	+856
Removed: bound forms, used only as the first part of compounds (កាម—, គុណ—)	38,781	−637
Removed: phrases containing spaces, digits or brackets (ព្រឹកសើច ល្ងាចយំ, កូវីដ១៩, ផលិតផលជាតិសរុប (ផ.ជ.ស.))	38,769	−12
Removed: malformed pronunciation (សម → សៈ-ម៉ៈ—)	38,768	−1
Removed: source error, a phrase that carries only its headword's pronunciation (ហាមាត់ចង្គ្រុង → ហា, រាជ្យសម្បត្តិ → រាច)	38,748	−20
Removed: source error, a truncated alternative reading (សរស្វតី → … ឬ ដី, វាចិកហារក → … ឬ កៈ)	38,740	−8
Duplicates removed	38,629	−111

All 28 rows removed as source errors were checked by hand.

Normalisation applied to the kept pairs:

• [ ] brackets and reading notes such as (ជើង ត) are removed from pronunciations.
• Trailing !, ? and ... are removed from words (interjections such as ផេះ! → ផេះ).
• Zero-width joiners and non-joiners (U+200C, U+200D) are removed from words and pronunciations.
• 6 source typos are fixed: : typed for ៈ (សុ-ប្រា-ត:-… → សុ-ប្រា-តៈ-…), stray ´ accents, doubled [[, and -- → -.

Deliberately kept: short readings caused by silent letters are correct and are not removed, e.g. ការណ៍ → កា, រាស្ត្រ → រាស, សួស្តិ៍ → សួស, ភ័ក្ដ្រ → ភ័ក. The truncation check ignores subscript consonants and letters silenced by ៍ / ៌ so that these words pass.

Intended uses

• Training and evaluating Khmer G2P / pronunciation models
• Text-to-speech front ends (word → pronunciation before synthesis)
• Syllabification: the - separators give syllable boundaries in the spoken form
• Reading aids and pronunciation lookup for learners

Limitations

• Dictionary words only. Covers standard vocabulary in the NCKL dictionary. There are no personal names, few place names, and no slang, digits, abbreviations or new loanwords.
• Standard (Phnom Penh) pronunciation as prescribed by NCKL, not regional or colloquial variants.
• Khmer-script respelling, not IPA. Most of the sound is explicit (register ៉ / ៊, short vowels ៈ, final consonants), but it is not a phonetic transcription.
• Words, not sentences. Each entry is one isolated word or compound. Pronunciation changes across word boundaries in running speech are not covered.
• Source errors that the automatic checks did not catch may remain.

Related datasets

• phonsobon/khmer-khmer-dictionary: the full dictionary (definitions, part of speech, examples), one row per sense, pronunciation as in the source.

License and attribution

The content comes from the Khmer Dictionary 2022 (វចនានុក្រមខ្មែរ ២០២២), © National Council of Khmer Language (NCKL), Royal Academy of Cambodia. Please credit NCKL when using this data, and check NCKL's terms before commercial use.