Dzongkha
Dzongkha is the national language of Bhutan [1]. AI support is new and small. Google Translate and a Bhutan government system can translate into Dzongkha, but there are no recent public quality scores.
Summary
| Where | Bhutan |
| Speakers | about 640,000 including second-language speakers (2013, via Wikipedia; weak source) [14] |
| Script | Tibetan script |
| AI resources | low |
| Google Translate | yes, added in June 2024 [2]; in the Cloud API as NMT only [3] |
| Microsoft Translator | no [4] |
| Speech to text | not in Whisper [8]; Meta Omnilingual ASR and a Bhutan MMS fine-tune [9] [1] |
| Best evidence on translation quality | FLORES-200, NLLB and ChatGPT, 2023 [5] |
Writing and text input
Dzongkha uses the Tibetan script, with 30 consonants and 4 vowel signs [1]. Bhutan has 23 languages, 21 of them indigenous [1].
In August 2023, Dzongkha OCR and a spelling and grammar checker were “planned future upgrades” [6]. We found no Dzongkha-specific OCR model. Tibetan OCR models (BDRC, OpenPecha) may work on printed Dzongkha because the script is the same, but this is not tested.
Data
- FineWeb-2 has 23,066 Dzongkha documents (32.15 MB) [10]. This is about seven times less than Tibetan in the same set.
- MADLAD-400 checked 1,899 Dzongkha documents. Its audit note says “maybe actually bo” (Tibetan) and “mainly buddhist vibes”, and Dzongkha is not in its final table [11]. So some web text labelled Dzongkha may be Tibetan.
- The Bhutan government MT system was trained on about 700,000 parallel sentences from the Dzongkha Development Commission (DCDD) [7]. A GovTech talk mentions a “10 million sentence parallel corpus” for an earlier version [1]. These two numbers do not agree, and we could not check which is right.
- The licenses of these Bhutan datasets are not stated.
Speech
- OpenAI Whisper does not list Dzongkha [8].
- Bhutan GovTech fine-tuned Meta MMS for Dzongkha. Speech to text (MMS-1B-All adapter, 7,385 training items) reached a word error rate (WER) of 0.373. Text to speech (mms-tts-dzo, 2,665 audio-text pairs) improved Mel Cepstral Distortion from 7 to 4.2 [1].
- Meta Omnilingual ASR (November 2025) reports a character error rate of 3.7 for Dzongkha, with 23.0 hours of training audio [9]. The test set is small, so this is a sign of support, not of real-world quality.
Machine translation
Support. Google Translate added Dzongkha in June 2024 [2]. In the Google Cloud API it is NMT only, not in Google’s Translation LLM [3]. Microsoft Translator does not support Dzongkha [4]. Meta NLLB-200 covers Dzongkha (dzo_Tibt) [12].
Bhutan government system. The College of Science and Technology (CST) launched the first Dzongkha MT system in August 2023, with “30 to 50 per cent” accuracy reported [6]. The current Dzongkha–English system is built on Meta NLLB and is “for testing and research only” [7]. A newer version based on NLLB-distilled-600M, Opus-MT and T5 is under test; the talk reports human evaluation but no scores [1].
Benchmark: FLORES-200 devtest, English to Dzongkha, chrF++ (higher is better).
| System | Score | Year | Source |
|---|---|---|---|
| NLLB-200 MoE | 34.7 | 2022 model, tested 2023 | [5] |
| NLLB-200 (likely 3.3B; copy of the official file, model not verified) | 32.6 | 2022 | [13] |
| ChatGPT (gpt-3.5) 5-shot | 15.9 | 2023 | [5] |
| ChatGPT (gpt-3.5) 0-shot | 7.7 | 2023 | [5] |
Google Translate was not scored for Dzongkha in that study [5]. These results are from 2023, and newer models may be much better. We found no public score for current frontier models. chrF++ scores can only be compared between systems in the same language and direction.
Risk. Because some web text labelled Dzongkha may be Tibetan [11], AI output may mix in Tibetan spelling and words. This is our inference, not a tested fact.
AI projects
- Dzongkha NLP by Bhutan GovTech (Data Science and AI Division), with DCDD and CST: MT, speech to text, text to speech and a chatbot prototype [7] [1].
- A Dzongkha LLM pilot on Llama 3.1 8B, Qwen2.5 7B and DeepSeek 7B. The talk gives the training data as both “1.2 lakh” (120,000) and “only around 20k”. An mT5 fine-tune reached 50% accuracy [1].
Recommendation
To translate into Dzongkha today, use Google Translate as a baseline. Also test the Bhutan government NLLB system (research use only) and one current frontier model. A native Dzongkha reviewer must check all output, especially for Tibetan words and spellings. Confidence is low: the only scores are from 2023, and the government’s own 2023 system reported 30 to 50 per cent accuracy.
Test first: FLORES+ devtest, English to Dzongkha, with chrF++, plus native review of 50 sentences per system. Then test a short Buddhist sample, such as prayers or a teaching summary. Dzongkha is not yet measured on the BuddhistAI leaderboard.
In the Directory
Projects and datasets in the Directory that list Dzongkha. This list is made from the Directory data. To add an entry, suggest it.
Projects (2)
- Dzongkha NLP (GovTech Bhutan) (GovTech Agency Bhutan with DCDD and CST): Dzongkha-English machine translation built on Meta NLLB (about 700,000 sentence pairs), Dzongkha speech to text and text to speech, and a chatbot prototype. For testing and research only.
- Dzongkha NLP (CST) (College of Science and Technology (Royal University of Bhutan) with DCDD): Web app with Dzongkha-English translation, Dzongkha speech to text and text to speech.
Datasets (0)
None listed yet.
Sources
- SANOG43: Dzongkha NLP (GovTech Bhutan and CST, slides), read 2026-11-17
- Google Translate Help: more than 100 new languages, read 2026-11-17
- Google Cloud Translation: Language support, read 2026-11-17
- Microsoft Learn: Translator language support, read 2026-11-17
- Robinson et al. 2023, arXiv 2309.07423, read 2026-11-17
- ThePrint/ANI: Dzongkha machine translation system launched in Bhutan, 2023-08-10, read 2026-11-17
- Bhutan GovTech Dzongkha NLP: background, read 2026-11-17
- OpenAI Whisper tokenizer.py (language list), read 2026-11-17
- Meta Omnilingual ASR, per-language results (7B LLM-ASR), read 2026-11-17
- FineWeb-2 README, read 2026-11-17
- MADLAD-400 README, read 2026-11-17
- NLLB-200 1.3B model card, read 2026-11-17
- Copy of the NLLB-200 metrics file (Mara4444/MSc_Thesis), read 2026-11-17
- Wikipedia, Dzongkha, read 2026-11-17.

BuddhistAI Knowledge base