Buy or build? (Spoiler: buy.)
A page on our own website recommending that you buy from us. We know how that sounds. But we’ve spent years building exactly this, so we can tell you honestly what it takes.
“Why don’t we just build this ourselves?”
Somewhere in a strategy meeting, someone has asked it. It’s a fair question. Here’s roughly what the answer involves:
- speech recognition engineers
- computational linguists and AI researchers
- UX designers who know what makes a 12-year-old want to talk to a bot
- specialists in second language acquisition
- a large collection of recordings of real learners (native speakers won’t do)
- a few years
- a budget your CFO will want to discuss at some length
And once it’s built, it needs maintaining, permanently, because AI keeps moving.
Or you integrate something that already exists, is used by around 100,000 pupils this school year, and is built by a team that has worked on nothing else for years. That’s us, obviously. This is our website. So let’s earn your trust rather than just claim it.
1. Speech recognition built for learners
This is the core of everything we do, so we’ll start here.
Off-the-shelf speech recognition from the big tech players was built for native speakers. It’s excellent at understanding a BBC newsreader. A 12-year-old from Antwerp in her third year of French is another story: roughly one word in four comes out wrong. You can’t give meaningful feedback on a transcript like that. You’d end up correcting things the pupil actually said right.
Our own models are trained specifically for second-language learners and stay below 10% word error rate. We keep refining them with the ESAT research group at KU Leuven. The system even selects the best model for each individual learner, based on their pronunciation profile, because accents vary enormously and one model for everyone simply doesn’t cut it.
On top of an accurate transcript, we analyse pronunciation at word and phoneme level, so learners know exactly which sounds need work. It’s not about penalising an accent (we all have one). It’s about being understood.
2. Grounded in research, and recently awarded
In July 2026 our adaptivity research won the Outstanding CALICO Journal Article Award, one of the main international prizes in computer-assisted language learning. Belgium didn’t bring the World Cup home that summer, but we did bring this. The paper, written with Itec (KU Leuven and imec), describes our AI learner model: it predicts how likely a learner is to complete a task, so each conversation offers the right challenge with the right amount of support.
Unlike a black-box chatbot, the model is transparent and theory-informed. And it isn’t stuck in a lab: it runs live in our exercises today.
We also work with ESAT (KU Leuven), IDLab (UGent) and UCLL, and our Head of Research and Learning Design keeps every exercise grounded in how people actually learn languages: task-based, with authentic interaction, meaningful output, corrective feedback and scaffolding. We could happily fill twenty pages on this. But you’re a publisher, not a review board, so we’ll keep it to this paragraph.
3. A conversation player with real depth
The conversation player is more than a chatbot with a pretty face, although it does offer a few faces: a WhatsApp-style chat, a video avatar, or both at once. The real depth is in what you can configure per exercise:
Built-in adaptivity
Beginners get simple tasks with plenty of support, advanced learners get challenged with minimal scaffolding. When frustration builds up, the exercise eases off by itself.
Flexible answer types
Speaking, writing, multiple choice, or speaking with written support in between.
Layered feedback
From short, gentle corrections to full grammatical and contextual feedback, matched to your target group and didactic choices.
Learner support
Hints when a learner goes quiet, a translation of any word on click, scaffolding in open conversations, and an adjustable bot speaking rate.
A motivating experience
Progress bar, visual rewards, sound effects, light and dark theme. Language practice should feel good, not like filling in a tax form.
Part of a full suite
Pronunciation exercises, interactive writing, more than 15 interactive formats, vocabulary and verb trainers: they all share the same feedback, dashboards and integration.
Every one of these features was requested by our publisher partners. We invested the time, the people and the research to build what the market asked for. It’s not a wish list, it’s what publishers use today.
4. Content creation, dashboards and privacy
A full conversation in 30 minutes
Your editors build a complete conversation exercise in about 30 minutes, starting from simple text input, without any coding or engineering background. Specialists who want fine-grained control (in our experience, that’s most publishers) can adjust every detail: structure, tasks, answer types, adaptivity and bot behaviour.
Dashboards that show what happened
Teachers can read every answer and replay every recording, or see at a glance which mistake the whole class keeps making. (There’s always one.)
Privacy without asterisks
Everything runs on our own servers, including the language models. That’s rare in our sector, and it matters more than most people realise. We decide 100% of what happens with pupil data: fully GDPR compliant, nothing leaves the EU, and no dependency on AI providers whose terms keep changing. When a school board asks where the recordings go, you can give a short, honest answer instead of a three-page disclaimer.
5. A team that knows publishing
Software engineers, a PhD in speech recognition, a PhD applied linguist heading research and learning design, and a business developer who used to be a language teacher and an educational publisher herself. Different backgrounds, one shared goal: helping people speak with confidence.
We’re not a tech company that discovered education last Tuesday. We’ve run more than 50 user tests, from Flemish primary schools to secondary schools in Amersfoort. When Noordhoff chose us, they told us what tipped the balance: the right expertise, our track record with Plantyn, and the fact that we came to the Netherlands to test in their classrooms. We show up. Literally.
So: build or buy?
| What you need | Build it yourself | Partner with Linguineo |
|---|---|---|
| Team | Specialists in speech, NLP, adaptivity and pedagogy, a rare combination on the job market | A team that has done nothing else for years |
| Speech models for learners | Years of collecting learner recordings and training models | Ready, below 10% word error rate |
| Time to first pupils | Years | Weeks: one HTML tag or the Moodle plug-in |
| Content | First build your own authoring tool | A content creation tool your editors learn in a day |
| Keeping up | Maintain models, conversation engine, adaptivity, feedback, dashboards and servers, indefinitely | We handle the AI, you handle content and pedagogy |
| Privacy | Design and run a GDPR-proof AI infrastructure | Already on our own EU servers, models included |
We won’t pretend we’re the only option on the market. But we will say this with confidence: nobody else combines learner-tuned speech recognition, award-winning adaptivity, full content control for publishers, around 100,000 pupils this school year and privacy-first infrastructure, all from one focused team.
That’s not marketing. That’s just what we’ve built. (All right, it’s a little bit of marketing. But it’s true.)
Ready to see it in action?
Request a demo and we’ll show you the conversation player with a scenario tailored to your method, along with the content creation tool, the dashboards and the integration. One conversation is all it takes.
Back to the overview
See everything our widgets add to your method, and which publishers already use them.