Welsh text-to-speech that gets every word right.
Fi is Diffian’s bilingual Welsh and English text-to-speech service for the organisations that serve Wales; say it “vee”, the Welsh word for “me”. Phone lines, appointment reminders, announcements, accessible information: wherever your service speaks, it can speak Welsh properly.
Generic speech engines fail in Welsh.
Welsh isn’t English with different words. Real service content mixes two languages in a single sentence, and the big cloud voices weren’t built for it. When we last checked, none of the major vendors would even train a custom voice that speaks Welsh.
Mixed-language sentences
Announcements, IVR menus and service messages switch between Welsh and English constantly. Single-language voices read one of them wrong every time.
Diffian: the language is detected word by word, so each word is read in its own language by one natural voice.
Place names
Llanelli, Pen-y-lan, Aberystwyth, Cwmbrân. English voices mangle Welsh place names, and Welsh speakers notice instantly.
Diffian: a curated lexicon of thousands of Welsh place names, built from national transport data.
Mutations
Welsh words change their first letters with grammar. Caerdydd becomes Gaerdydd or Nghaerdydd depending on context, and engines that only know dictionary forms get real sentences wrong.
Diffian: pronunciation entries cover mutated forms, with per-word overrides where rules fall short.
Hear it for yourself.
Every clip was generated by Carys, our own neural voice, on Fi, the same engine your service would use. The line underneath each sentence is the engine’s actual phonetic output.
Need a bespoke voice or a full announcement pack? Talk to us.
Our own voice. Our own engine. Built for Welsh from the ground up.
Fi is the engine and Carys is its voice. When we found that no cloud vendor could give us a custom voice that speaks Welsh, we trained our own. Carys is one neural voice that speaks both languages, so her English says Welsh place names properly and her Welsh flows like a first language. And because the voice belongs to us, it can belong to your service: no per-character licensing and no vendor lock-in.
Carys, a voice you can own
A custom-trained neural voice producing 24 kHz audio, efficient enough to run on ordinary CPU servers. Self-host it on your own infrastructure or consume it as a managed API. Your content never has to leave your estate.
Bilingual by design
Per-word language detection handles code-switching automatically, with explicit control when you need it. One request, one voice, both languages, correctly.
Fine control
An SSML subset covers pauses, prosody, substitutions and exact phoneme overrides, plus Welsh and English number expansion. Your Welsh language officer’s corrections can be applied precisely and regenerated in seconds.
Delivery that fits your hardware
A simple HTTP API for live generation, or pre-generated offline audio bundles for annunciators, queue systems and kiosks. GPS-triggered on-vehicle playback needs no connectivity at all.
Or make it your own voice.
Carys is ready today. But the same pipeline that trained her can train a voice that belongs to your organisation alone: a distinct, consistent identity across every automated channel, in both languages. That’s the promise in the name: Fi is Welsh for “me”.
Start with Carys
A production-ready bilingual voice you can deploy now, under your control: self-hosted on your infrastructure or as a managed API, with no per-character meter and no vendor lock-in.
Or train a voice that is yours alone
We build a bespoke neural voice from recordings of a speaker you choose. It speaks Welsh and English the way Carys does, mutations, place names and mixed sentences included, but it sounds like your organisation and belongs to it. No cloud vendor offered this for Welsh when we last checked, in May 2026.
First test: every bus stop in Wales.
The PSV Accessible Information Regulations require audible next-stop announcements on local bus services from October 2026. In Wales, that means saying every settlement, street and interchange name correctly, in both languages. Fi was built to solve exactly that, then run across the entire national timetable dataset.
Anywhere your organisation speaks, it can speak Welsh well.
Built for the organisations that serve Wales, wherever a service talks to the public.
Contact centres & IVR
Few organisations can staff every phone line and automated touchpoint with Welsh speakers. Give your IVR menus, queue messages and automated lines a fluent Welsh-first voice, without recruiting for it.
Health boards
Appointment reminders and automated calls, outpatient queue calling, spoken patient information and Welsh-first telephone services. Patient-derived text never has to leave your network: CPU-only self-hosting materially simplifies the DPIA.
Public transport
The Public Service Vehicles (Accessible Information) Regulations 2023 require audible next-stop announcements on local bus services, with deadlines from October 2026. Diffian generates complete bilingual announcement sets packaged for standard on-vehicle annunciator hardware.
Councils & public bodies
Councils, national parks and museums: tannoy announcements, wayfinding, queue systems and digital services, generated in both languages from a single source, with every place name in your area said properly.
Digital services
Read-aloud on public-facing web content and apps, so a Welsh speaker gets Welsh audio rather than an English voice attempting it.
Built around the rules services in Wales work to.
Welsh Language Standards
Welsh-first service delivery for public bodies in Wales, including telephone and automated services. Diffian generates Welsh and English from one workflow, with Welsh leading, the way the Standards expect it.
Accessible information
Spoken versions of written patient and service information, for people with sight loss or who find printed information difficult, in the recipient’s language of choice.
PSV Accessible Information Regulations 2023
For local bus services: audible and visible route information, with deadlines from October 2026. Our announcement templates and offline bundles were designed for exactly this requirement.
No supplier can certify you as compliant, and we won’t claim to. What we provide is the generation capability and an audit trail: deterministic output you can verify by hash, and inspectable phonetics. Editorial sign-off on your content remains yours. Our tooling is built to make that sign-off fast, not to bypass it.
Talk to us.
Tell us what you’re working on and we’ll come back to you, usually the same day. No commitment, no procurement forms.
Prefer email? hello@diffian.com