Welsh to English: AI Translation Comparison
Welsh to English: AI Translation Comparison
Welsh (Cymraeg) is spoken by approximately 880,000 people, primarily in Wales but also in diaspora communities in England, Patagonia (Argentina), and North America. As a Brythonic Celtic language, Welsh exhibits several features that challenge AI translation systems: verb-subject-object (VSO) word order (the default sentence structure places the verb first), an elaborate system of initial consonant mutations (soft, nasal, and aspirate) that alter word beginnings based on grammatical context, and a vigesimal (base-20) counting system in traditional forms. Welsh also uses inflected prepositions, where prepositions conjugate like verbs based on person and number. The language has experienced a significant revival since the Welsh Language Act of 1993 and the establishment of the Senedd (Welsh Parliament), which operates bilingually. Translation demand is driven by Welsh government services, bilingual public signage requirements, education, broadcasting (S4C), literature, and the growing Welsh-medium digital ecosystem.
This comparison evaluates five leading AI translation systems on Welsh-to-English accuracy, naturalness, and suitability for different use cases.
Translation comparisons are based on automated metrics and editorial evaluation. Quality varies by language pair and content type.
Accuracy Comparison Table
| System | BLEU Score | COMET Score | Editorial Rating (1-10) | Best For |
|---|---|---|---|---|
| Google Translate | 27.8 | 0.785 | 6.6 | General-purpose, everyday translation |
| DeepL | 23.5 | 0.752 | 5.7 | Basic document translation |
| GPT-4 | 30.2 | 0.804 | 7.2 | Complex grammar, literary and formal content |
| Claude | 28.4 | 0.791 | 6.8 | Government documents, long-form content |
| NLLB-200 | 28.9 | 0.795 | 6.9 | Free, self-hosted, consistent quality |
Translation Quality Metrics: BLEU, COMET, and Human Evaluation Explained
Example Translations
Formal Business Email
Source: “Annwyl Mr. Davies, Ysgrifennaf atoch i gadarnhau bod y cytundeb masnachol rhwng ein dwy gwmni wedi ei arwyddo yn swyddogol ddydd Llun diwethaf. Edrychwn ymlaen at gydweithio llwyddiannus.”
| System | Translation |
|---|---|
| Dear Mr. Davies, I write to you to confirm that the commercial agreement between our two companies has been officially signed last Monday. We look forward to a successful collaboration. | |
| DeepL | Dear Mr. Davies, I am writing to you to confirm that the business agreement between our two companies was officially signed last Monday. We look forward to successful cooperation. |
| GPT-4 | Dear Mr. Davies, I am writing to confirm that the commercial agreement between our two companies was officially signed last Monday. We look forward to a successful collaboration. |
| Claude | Dear Mr. Davies, I write to you to confirm that the commercial agreement between our two companies has been officially signed last Monday. We look forward to successful collaboration. |
| NLLB-200 | Dear Mr. Davies, I write to you to confirm that the commercial agreement between our two companies was officially signed last Monday. We look forward to successful collaboration. |
Assessment: All systems handle this formal content well, benefiting from the substantial Welsh-English parallel corpus available through the Welsh Government’s bilingual publishing requirements. GPT-4 produces the most natural English by using “I am writing to confirm” (without the redundant “to you”) and correctly choosing past simple “was officially signed” over the awkward “has been officially signed last Monday” that Google and Claude produce. The tense error in Google and Claude’s versions is notable — “has been” with “last Monday” is incorrect in standard English.
Casual Conversation
Source: “Shwmae, sut wyt ti? O’n i’n meddwl mynd lawr i’r traeth pnawn ‘ma, ma’r tywydd yn fendigedig. Ti ffansi dod? Allwn ni gael hufen ia wedyn.”
| System | Translation |
|---|---|
| Hello, how are you? I was thinking of going down to the beach this afternoon, the weather is wonderful. Do you fancy coming? We can have ice cream after. | |
| DeepL | Hello, how are you? I was thinking of going to the beach this afternoon, the weather is wonderful. Do you fancy coming? We could have ice cream afterwards. |
| GPT-4 | Hiya, how are you? I was thinking of heading down to the beach this afternoon — the weather’s absolutely gorgeous. Fancy coming along? We could grab an ice cream after. |
| Claude | Hello, how are you? I was thinking of going down to the beach this afternoon, the weather is wonderful. Do you fancy coming? We could have ice cream afterwards. |
| NLLB-200 | Hello, how are you? I was thinking of going to the beach this afternoon, the weather is wonderful. You fancy coming? We can have ice cream then. |
Assessment: GPT-4 captures the casual Welsh tone best with “Hiya” for “Shwmae” (a distinctly informal Welsh greeting), “heading down,” “gorgeous,” and “grab an ice cream” — natural British English casual register that matches the Welsh source. Notably, “fancy coming” without “do you” is more colloquially authentic. Other systems produce grammatically correct but comparatively stiff translations. The Welsh soft mutation “fendigedig” (from “bendigedig”) is correctly handled by all systems. How AI Translation Works: Neural Machine Translation Explained
Technical Content
Source: “Mae’r system ynni adnewyddadwy yn defnyddio tyrbinau gwynt ar y mor ynghyd a phaneli solar ar y tir i gynhyrchu trydan ar gyfer y grid cenedlaethol, gan leihau dibyniaeth ar danwyddau ffosil.”
| System | Translation |
|---|---|
| The renewable energy system uses offshore wind turbines along with solar panels on land to produce electricity for the national grid, reducing dependence on fossil fuels. | |
| DeepL | The renewable energy system uses offshore wind turbines and onshore solar panels to generate electricity for the national grid, reducing reliance on fossil fuels. |
| GPT-4 | The renewable energy system employs offshore wind turbines in conjunction with onshore solar panels to generate electricity for the national grid, thereby reducing dependence on fossil fuels. |
| Claude | The renewable energy system uses wind turbines at sea along with solar panels on land to generate electricity for the national grid, reducing dependence on fossil fuels. |
| NLLB-200 | The renewable energy system uses wind turbines at sea and solar panels on land to produce electricity for the national grid, reducing dependence on fossil fuels. |
Assessment: GPT-4 produces the most polished technical English with “employs,” “in conjunction with,” and “thereby reducing,” which are appropriate for energy industry documentation. DeepL efficiently renders “ar y mor” and “ar y tir” as “offshore” and “onshore” — standard energy sector terminology. Claude and NLLB-200 translate more literally with “at sea” and “on land,” which is accurate but less industry-standard. The Welsh mutation “phaneli” (nasal mutation of “paneli” after “a”) is correctly resolved by all systems.
Strengths and Weaknesses
Google Translate
Strengths: Decent baseline quality. Benefits from Welsh Government bilingual corpora. Free and accessible. Weaknesses: Occasional tense errors. Literal approach to mutations and idioms. Limited register adaptation.
DeepL
Strengths: Efficient, clean English output. Good technical vocabulary. Weaknesses: More limited Welsh training data. Misses some Welsh-specific idioms. Weaker on colloquial registers.
GPT-4
Strengths: Best contextual understanding. Excellent register adaptation. Handles mutations and VSO word order reliably. Produces natural British English. Weaknesses: Higher cost. May occasionally over-localize to British English. Slower for bulk processing.
Claude
Strengths: Reliable for long government documents. Consistent quality. Good formal register. Weaknesses: Occasional tense inconsistencies. Less creative with casual content. Sometimes too literal.
NLLB-200
Strengths: Free and self-hostable. Strong baseline quality. Good mutation handling. Weaknesses: Literal translations of idioms. Drops some natural phrasing. No register adaptation.
Recommendations
| Use Case | Recommended System |
|---|---|
| Quick personal translation | Google Translate (free) |
| Welsh Government documents | Claude or GPT-4 |
| Literary translation | GPT-4 with human review |
| Bilingual signage and public communications | GPT-4 or Claude |
| Broadcasting subtitles (S4C) | GPT-4 |
| High-volume processing | NLLB-200 (self-hosted) |
| Education materials | Claude |
Best Translation AI in 2026: Complete Model Comparison
Key Takeaways
- GPT-4 leads for Welsh-to-English translation with the strongest handling of VSO word order restructuring, initial consonant mutations, and register adaptation to natural British English.
- Welsh benefits from unusually strong institutional support for a language of its size: the Welsh Government’s bilingual publishing mandate creates substantial parallel corpora that boost all systems’ baseline quality.
- NLLB-200 provides a competitive free alternative, performing especially well on formal and technical content where its literal approach is an asset rather than a limitation.
- The primary differentiator between systems is casual and literary Welsh, where idiomatic expressions, dialect variation, and register sensitivity separate contextual models like GPT-4 from rule-following systems.
Next Steps
- Try it yourself: Compare these systems on your own text in the Translation AI Playground: Compare Models Side-by-Side.
- Check the leaderboard: Browse our full Translation Accuracy Leaderboard by Language Pair.
- Understand the metrics: Learn what BLEU and COMET scores mean in Translation Quality Metrics.
- Explore rare languages: Read Best AI Translation for Rare and Low-Resource Languages.