✎ Edit this page

Sanskrit

Buddhist Sanskrit is the language of many Mahāyāna sūtras and śāstras, and of manuscripts kept in Nepal. There is good data for translation from Sanskrit into English, but every system was weak at writing Sanskrit in the tests we found.

Summary

   
Where India and Nepal; scholars worldwide
Speakers Google described Sanskrit as “used by about 20,000 people in India” (2022) [1]
Script Devanagari; Roman (IAST) in scholarship
AI resources low
Google Translate yes, since 2022-05-11 [1]; NMT only in the Cloud API [2]
Microsoft Translator no [3]
Speech to text Whisper lists Sanskrit [4]; not in Meta Omnilingual ASR [5]
Best evidence on translation quality FLORES-200, English to Sanskrit, 2023 [6]

Writing and text input

Sanskrit is written in Devanagari in India and Nepal. Scholars often use Roman letters with diacritics. Tools for sandhi splitting and lemmas (ByT5 Sanskrit analyzers, chronbmm data) and a Buddhist Sanskrit BERT model are listed in the Directory below. We did not verify OCR support for Sanskrit in this research.

Data

Almost all parallel data and tests go from Sanskrit into English. Very little is made for English into Sanskrit.

Speech

Machine translation

Support. Google Translate added Sanskrit on 2022-05-11 [1]. In the Google Cloud API it is NMT only [2]. Microsoft Translator does not support Sanskrit [3]. AI4Bharat’s IndicTrans2 covers Sanskrit (san_Deva, about 278,000 training pairs) and aims at permissive licenses [10]. Meta NLLB-200 also covers Sanskrit [6].

Benchmark: FLORES-200 devtest, English to Sanskrit, chrF++ (higher is better), 2023 [6].

System Score
Google Translate 30.3
NLLB-200 MoE 26.1
ChatGPT (gpt-3.5) 5-shot 22.6
ChatGPT (gpt-3.5) 0-shot 21.8

A copy of the NLLB-200 metrics file (model not verified, likely 3.3B) gives 25.6 [11]. All systems were weak. These are 2023 results, and newer models may be better. chrF++ can only be compared between systems in the same language and direction.

Benchmark: Sāmayik, English to Sanskrit, chrF. Best in-domain 46.61 (fine-tuned mBART); best out-of-domain 26.7 (fine-tuned IndicTrans) [9].

Benchmark (opposite direction): Mitrasaṃgraha test set, Sanskrit to English, chrF / BLEURT / GEMBA (higher is better) [7].

System chrF BLEURT GEMBA
Claude 3.7 Sonnet with retrieval (RAG) 38.83 57.54 89.32
Fine-tuned Gemma 2 9B 38.36 55.10 83.34
Gemini 2.0 Flash, zero-shot 34.24 53.87 87.04
Google Translate 28.28 47.42 56.09

These scores show how well models read Sanskrit. They do not show how well models write Sanskrit.

AI projects

Recommendation

To translate into Sanskrit today, use a current frontier model with retrieval (give it similar Sanskrit passages and a glossary). Use Google Translate and IndicTrans2 as baselines. A Sanskrit scholar must review all output. Confidence is low: the only into-Sanskrit scores are from 2023 and are low for every system.

Test first: FLORES+ devtest, English to Sanskrit, with chrF++, plus a Buddhist sample. For the sample, reverse 50 to 100 pairs from the Mitrasaṃgraha test set (English in, Sanskrit out) and ask a scholar to review 20 items. Check the license before you share any derived data. Sanskrit is not yet measured on the BuddhistAI leaderboard.

In the Directory

Projects and datasets in the Directory that list Sanskrit. This list is made from the Directory data. To add an entry, suggest it.

Projects (16)

Datasets (8)

Sources

  1. Google blog: 24 new languages, 2022-05-11, read 2026-11-17
  2. Google Cloud Translation: Language support, read 2026-11-17
  3. Microsoft Learn: Translator language support, read 2026-11-17
  4. OpenAI Whisper tokenizer.py (language list), read 2026-11-17
  5. Meta Omnilingual ASR language IDs, read 2026-11-17
  6. Robinson et al. 2023, arXiv 2309.07423, read 2026-11-17
  7. Nehrdich et al., Mitrasaṃgraha, arXiv 2601.07314, 2026-01-12, read 2026-11-17
  8. Nehrdich and Keutzer, MITRA, arXiv 2601.06400, 2026-01-10, read 2026-11-17
  9. Sāmayik, arXiv 2305.14004 (Hugging Face papers page), read 2026-11-17
  10. IndicTrans2, arXiv 2305.16307, read 2026-11-17
  11. Copy of the NLLB-200 metrics file (Mara4444/MSc_Thesis), read 2026-11-17

✎ Edit this page · History · Discuss