How Diffian speaks Welsh.
Most speech engines treat text as one language at a time. Welsh public-service content doesn’t work that way, so we built an engine that reads the way Wales actually writes and speaks.
A voice we trained ourselves
Diffian’s voice, Bronwen, is a custom neural voice trained on our own recorded corpus. We went this way for a simple reason: we verified that none of the major cloud vendors can train a custom voice that speaks Welsh. Welsh simply isn’t in their custom-voice programmes. If you want a voice that is both distinctively yours and genuinely Welsh, it has to be built, not bought.
Because the model is ours end to end:
- It produces natural 24 kHz audio and runs comfortably on CPU-only servers. A warm request generates a clip in well under a second, with no GPU bill attached.
- There is no per-character licensing and no usage meter owned by someone else.
- It can be deployed wherever your data governance requires, including entirely inside your own network.
The same training pipeline that produced Bronwen can produce a bespoke voice for your organisation.
Pronunciation engineering, not just synthesis
The synthesis model is only half the work. The other half is deciding, for every single word, how it should be said. Diffian’s front end works per word, and exposes its decisions as phonetic transcription you can inspect and correct.
Per-word language detection
Every word is classified before it reaches the voice, so Welsh place names inside English sentences are read the Welsh way. This is the engine’s actual output:
When automatic detection isn’t what you want, language can be marked explicitly per word or per phrase with SSML tags, so editorial control always wins.
A national place-name lexicon
We mined the national transport dataset for Wales to build a curated lexicon of thousands of Welsh place names, cross-checked against an English dictionary so names like Penny, Lloyd or Trevor aren’t misread as Welsh. Hyphenated compounds (Pen-y-lan) and the Welsh article are handled as units.
Mutation-aware pronunciation
Welsh initial consonant mutations change the start of words in ordinary sentences. Caerdydd appears as Gaerdydd or Nghaerdydd depending on grammar, and our pronunciation overrides cover mutated forms explicitly, so real sentences sound right, not just dictionary citations.
Exact overrides when you need them
For the long tail, every word can carry an exact phonetic spelling, and the engine exposes its phoneme output so corrections are precise rather than guesswork. Combined with SSML support for pauses, prosody, substitutions and bilingual number expansion, a Welsh language officer’s review can be applied word by word and regenerated in seconds.
Three ways to deploy
Same engine, same voice, delivered the way your infrastructure and governance need it.
Managed API
A simple HTTP endpoint: send text or SSML, receive audio. Generated clips are cached, so repeated content returns in milliseconds. We run it, you integrate it.
Self-hosted
The full engine runs on your own servers. CPU-only, no GPU fleet, no external calls. The right fit where content or data can’t leave your estate.
Offline audio bundles
For announcement hardware, we pre-generate complete audio sets and package them for standard annunciator systems. Vehicles sync at the depot and play offline, GPS-triggered, with no on-vehicle connectivity required.
Send us your hardest sentence.
Place names, mutations, mixed languages. We’ll send it back spoken, usually the same day.
Talk to us