Word Exchange Plaza started as a tool for ESL conversation practice. It is becoming something larger. We are pivoting Word Exchange Plaza into a Duolingo-style language learning platform that supports many languages, including Hindi, Arabic, French, Norwegian, and a growing list beyond that. The bigger change is who builds the courses. Instead of a single editorial team shipping the same path to every learner, Word Exchange Plaza is becoming a platform where teachers create their own courses, run their own ads inside them, and produce studio-quality audio in minutes with an admin interface built on top of ElevenLabs. This article walks through the reasoning behind the pivot, the technical architecture that supports it, and what it means for learners, teachers, and developers interested in building the next generation of language learning tools.
Why Pivot From ESL to a Multi-Language Platform
The original Word Exchange Plaza was focused on English as a second language. The product worked well, but it bumped into the same ceiling that every single-language tool eventually hits. Learners who finished an English module would ask, reasonably, where the French module was. Or the Hindi module. Or the Arabic module. Building each of those ourselves would take years and a content team we do not have. More importantly, even if we did build them, every learner on the planet would get a path designed by a small team in Ottawa. That is not how language teaching actually works at its best.
The best language teachers are local. They know which idioms a specific cohort will hear in the street. They know which mistakes their students always make because of their first language. A teacher in Mumbai building a Hindi course for English speakers will make different choices than one in Delhi building the same course for Arabic speakers. A French teacher in Quebec writes different example sentences than a French teacher in Lyon. The platform that wins long term is not the one with the most languages. It is the one that gives the people who already know how to teach those languages the best tools to do it.
That insight pushed us toward a Duolingo-style structure for the lessons themselves — short, gamified, repeatable units — but with a fundamentally different posture toward content. Duolingo writes its own courses. Word Exchange Plaza hosts other people's.
What the New Platform Looks Like for Learners
The learner-facing experience is the part that should feel the most familiar. Short lessons, vocabulary drills, listening comprehension, speaking practice, and a streak you can lose at midnight if you are not careful. The pedagogical patterns Duolingo and other modern language apps have validated are well understood. There is no reason to reinvent them. Where Word Exchange Plaza differs is in the catalog and the voice.
When a learner opens the app, they can browse courses by language and by teacher. A French course taught by a Parisian high school instructor reads and sounds differently than a French course taught by a Montreal-based francisation tutor. A Hindi course aimed at heritage learners — second-generation kids who heard the language at home but never studied it formally — looks different from a Hindi course written for an absolute beginner with no exposure. Learners pick the teacher whose approach matches their goal. Some learners take multiple courses in the same language for the variety.
Audio is everywhere. Every word, every sentence, every example dialogue has narration. The audio is not a recording of the teacher's voice unless the teacher wants it to be. More often it is generated, in studio quality, through the integrated ElevenLabs pipeline described later in this article. This matters because consistent, clear audio is one of the strongest predictors of whether a learner sticks with a language course. Patchy audio kills retention faster than almost any other product flaw.
Hindi, Arabic, French, Norwegian, and Beyond
Our launch catalog is intentionally diverse. Hindi, Arabic, French, and Norwegian are the first four languages we are seeding with teacher partners, and each was chosen for a different reason.
Hindi is the largest language by absolute speaker count that is consistently underserved by mainstream language apps. Mainstream apps tend to ship a single Hindi course, often light on script support and heavy on tourist phrases. We want a catalog that includes courses for heritage learners, courses for tourists, courses for business travelers, and courses aimed specifically at people learning Devanagari from scratch.
Arabic is more complex because there is no single Arabic. Modern Standard Arabic is the language of news and formal writing, but no one speaks it as a native dialect. Egyptian Arabic, Levantine Arabic, Gulf Arabic, and Maghrebi Arabic are mutually less intelligible than people often assume. A platform that lets a Cairo-based teacher publish an Egyptian Arabic course alongside a Modern Standard Arabic course from a university lecturer solves a real problem that monolithic apps cannot.
French was included because the demand never softens. It is the second most studied language in North American schools, the working language of large parts of Africa, and an enormously useful business language. The trick with French is voice and register. A platform that lets teachers offer Quebec French, European French, and West African French as distinct paths is more honest about the language than one that pretends there is only one version.
Norwegian rounds out the launch because of a different kind of opportunity. The total addressable market is small, but the existing alternatives are thin. Mainstream apps have a single Norwegian course or none at all. A teacher in Oslo who already builds materials for adult learners has a real audience waiting for a better tool. Smaller languages are where teacher-led platforms can outperform the giants the most clearly.
Beyond the launch four, the architecture does not care which language is added next. Spanish, Mandarin, Japanese, Portuguese, Tagalog, Swahili — all of them work the same way on the platform. The constraint is finding teachers who want to ship a course, not the technology that hosts it.
Teachers Create Their Own Courses
The single most important design decision in the new Word Exchange Plaza is that the people who know how to teach a language are the ones who build the courses. The admin interface is a course authoring tool, not a content management system. Teachers do not just upload a list of words. They construct lessons.
A lesson in Word Exchange Plaza is a sequence of exercises, each tied to a small set of vocabulary or grammar points. Teachers can mix and match exercise types: multiple choice, listening, type-what-you-hear, speak-the-sentence, fill in the blank, match pairs, and a few others. They write the source content — the sentences in the target language and the translations — and they choose which exercises pull from which vocabulary sets. The platform handles the spaced repetition scheduling, the streak logic, the progress tracking, and the audio.
For teachers who already have curriculum materials, importing is straightforward. A CSV of vocabulary pairs can become the seed of a course in minutes. From there the teacher iterates: adds example sentences, groups related vocabulary into thematic units, drafts the dialogues that will anchor each section, and refines the difficulty curve. The work that takes a long time is the pedagogical work — choosing which words to teach in what order, writing examples that feel natural rather than translated — and the platform stays out of the way of that work.
The technical patterns powering this admin live in the same family as the ones we covered in building AI-powered web applications: a small set of server-rendered authoring pages backed by a streaming API for the bits that are AI-assisted. Where AI helps is exactly where teachers want help — generating candidate example sentences they can edit rather than write from scratch, suggesting distractors for multiple choice questions, flagging vocabulary that may be too advanced for the lesson's stated level. AI never publishes anything. The teacher is always in the editor's seat.
Teachers Serve Their Own Ads
The other unusual choice we made is letting teachers run their own ads inside their own courses. This is a meaningful departure from the platform-mediated ad market that most apps default to. The reasoning is simple: teachers are usually the best judge of what their students actually want to see.
A teacher whose Norwegian course attracts heritage learners can run a tasteful ad for a Norwegian-Canadian cultural society. A French teacher in Montreal can promote a local conversation group that meets in a café on Saint-Denis. A Hindi teacher running a course aimed at heritage learners might promote a community festival, a local restaurant, or their own paid one-on-one coaching. These ads are more contextual than any algorithmic ad market would ever produce, because the teacher already knows the audience.
Operationally, the ad slots are inventory the teacher controls. They can choose to leave them empty, fill them with self-promotion, accept third-party sponsorships, or sell them to local businesses. Word Exchange Plaza provides the slot, the analytics, and the policy guardrails. Teachers handle the rest. The platform takes a small share of paid placements, and the rest flows to the teacher.
This model only works because we trust the teachers. The flip side is that we have a clear standards document about what ads cannot appear, and a fast takedown process when something violates it. The platform is responsible for the floor, not the ceiling. Teachers get a lot of latitude, but not infinite latitude.
The ElevenLabs Audio Pipeline
The piece of the admin that gets the most enthusiastic reactions from teachers is the audio creation system. Generating high-quality, multilingual audio used to be the most painful part of building a language course. You either hired voice actors, which is expensive and slow, or you recorded the audio yourself in a closet with a USB microphone and hoped your accent was clean enough not to teach the wrong pronunciation. Neither path was good for ambitious courses.
ElevenLabs changed the economics of voice generation. The voices are good enough that learners cannot reliably tell they are synthesized. The accent control is precise enough to differentiate Quebec French from Parisian French, or Egyptian Arabic from Levantine Arabic. The cost per minute of generated audio is small enough that a teacher can iterate freely on a lesson without worrying about budget.
The integration in the Word Exchange Plaza admin is intentionally streamlined. When a teacher writes a sentence in a lesson, the audio for that sentence is generated automatically in the background. The teacher picks a voice from a curated set per language — male and female options, regional variants, slow and natural-pace pairs — and the audio is queued, generated, and cached. If a teacher edits the sentence, the audio regenerates. If they prefer a different voice, every sentence in the lesson can be re-narrated with one click.
Under the hood, the pipeline batches audio generation requests intelligently. New sentences in a draft lesson are generated only when the teacher confirms the text. Cached audio is reused whenever a sentence is repeated across exercises. Voice changes trigger only the affected sentences to regenerate, not the entire lesson. These are the same kinds of latency and cost optimizations we have applied in our other AI projects, including patterns described in our piece on how real-time language translation works, where streaming and incremental generation are central to keeping the experience responsive.
The result is that a teacher can publish a fully narrated lesson within an hour of opening the admin for the first time. The bottleneck is the teaching, not the production.
Why This Architecture Matters for Teachers
Stepping back from the individual features, the underlying claim is that the next generation of language learning platforms will be teacher-led the same way that platforms like YouTube, Substack, and Patreon are creator-led. The pattern is consistent: build the infrastructure, host the audience, and get out of the way of the people who actually make the thing.
For teachers, this changes the economics of building a course. The cost of producing studio-quality, multi-hour language content goes from tens of thousands of dollars to roughly the cost of a few months of platform fees. The time to launch goes from a year to a weekend if the teacher is motivated. And the revenue model goes from "hope a publisher buys your textbook" to "earn from learners, sponsorships, and your own promoted offers, week by week."
This does not mean any teacher can publish a great course. Course design is hard, and not everyone who knows a language can teach it well. The platform amplifies what teachers can do. It does not replace pedagogical skill. But the teachers who do have that skill no longer need to be employed by a publisher or a school to reach an audience.
What This Means for the Original ESL Audience
One of the questions we got most often when we started talking about the pivot was whether ESL learners — the original Word Exchange Plaza audience — would lose access to what they had been using. The answer is no. ESL courses become one slice of the new catalog rather than the entire product. The same patterns we built for AI-powered ESL conversation practice, which we wrote about in how AI is transforming ESL learning, still live inside Word Exchange Plaza. They are now joined by Hindi, Arabic, French, Norwegian, and whatever languages teachers ship next.
If anything, the pivot improves the ESL experience. Heritage Spanish speakers learning English get to learn from teachers who explicitly designed for them. Arabic speakers learning English can choose between teachers who teach contrastively against Modern Standard Arabic and teachers who teach against specific dialects. The same teacher-led logic that opens the platform to Hindi or Norwegian also makes English courses more specific and more useful for the learner.
Technical Stack and Lessons From the Pivot
For developers curious about how a pivot of this scope gets executed without throwing away the previous product, a few notes on the technical side. The core lesson is that the original Word Exchange Plaza was built with separation of concerns that we did not fully appreciate at the time. The exercise engine, the spaced repetition scheduler, the audio pipeline, and the progress tracking were already independent modules. What changed in the pivot was the content authoring layer and the language metadata. The runtime — the part learners actually see — barely needed to move.
The conversational AI components, which we have written about extensively in our guide to building conversational AI agents, did need updating. Conversation practice that worked for ESL did not directly transfer to languages with different script systems or different morphological complexity. Hindi conjugation patterns, Arabic root-and-pattern morphology, and Norwegian noun gender all required adjustments in how the AI assesses learner responses. We rebuilt those assessment modules per language family rather than per language, which kept the work tractable.
The infrastructure side has been deliberately boring. The streaming and media plumbing is closer to what we have used for live video work — patterns from budget live streaming on a $5 VPS apply here as well, because cached, ElevenLabs-generated audio behaves a lot like cached video segments. Storage costs are modest. Compute costs are dominated by ElevenLabs generation, which is bounded because each unique sentence is generated once and cached forever.
The single biggest engineering decision was making the admin a first-class product rather than an afterthought. Teachers spend more time in the admin than learners spend in the app, in terms of hours per active user. If the authoring experience is bad, teachers leave, and without teachers there is no platform. We treat every admin improvement as a learner-facing improvement, one layer removed.
What Comes Next
The next milestones are honest. Onboard more teachers in the launch languages and prove the model works at scale per teacher. Add the languages that teachers ask us to add, in the order they ask. Improve the assessment modules for languages we ship next, especially those with non-Latin scripts and complex morphology. Open the platform to teacher-led group classes alongside self-paced courses, because some teachers want to teach live and the technology to host them well already exists. And keep the admin fast, because every minute saved in lesson production is a minute spent on actual teaching.
The longer-term bet is that language learning is moving from a handful of giant general-purpose apps to a much larger ecosystem of specialized, teacher-led platforms. Word Exchange Plaza is one bet on what that ecosystem looks like. It is teacher-led, multi-language by default, AI-assisted but not AI-driven, and built on infrastructure cheap enough that the economics work even for small languages.
If you teach a language and you have wanted to build a course of your own, this is the platform we built for you. If you are learning a language and you want a course that matches your background rather than the average background of every learner in the world, this is the catalog we are building for you. And if you are a developer interested in this kind of work, the lessons we have learned along the way are the kind of thing we keep writing about on this blog and on the projects page.