Duolingo Max Babbel Speak ELSA Speak AI features

Four products, one retrofit problem

Duolingo, Babbel and ELSA Speak were all finished products before a machine could hold a spoken conversation. Speak was younger but built its business on a bounded repeat-and-score interaction that also predates open dialogue. Each of them has since attached a generative speaking capability to a loop that was designed without one, and the attachment is the most informative thing on any of their marketing pages.

Retrofits are not inherently worse than rebuilds. A retrofit inherits distribution, a content library, a brand and a payment relationship, which is an enormous head start. What it cannot inherit is permission to change the thing everyone already loves, and that constraint is what shapes the result. The new capability has to be placed somewhere it will not disturb the existing metric, which usually means beside the core loop rather than inside it.

This post reads the four attachments as engineering artefacts: where the feature sits, what it is allowed to touch, and what the placement costs. Klepha examines the same four feature sets for how assistants describe them, which is the retrieval question. This one is the product question.

Where a feature lives is the tell

There is a rule of thumb that survives contact with almost every consumer app: a capability that lives in its own tab was built by a team that was not allowed to modify the main flow. A capability woven into the main flow was built by the team that owns the main flow. You can usually read the reporting lines off the tab bar.

Applied here, the pattern is unusually clean. The generative speaking features on Duolingo arrive as an upgrade SKU with its own entry points sitting alongside the lesson tree rather than rewriting it. Babbel places its speaking practice adjacent to the course rather than as the spine of it, because the course is the product and the course was written by human editors. ELSA Speak keeps its conversation surface separate from the pronunciation scoring engine that made its reputation. Speak comes closest to integration and still fences open dialogue behind a tier.

None of that is laziness. It is what happens when a new capability has a materially different cost curve and a materially different failure mode from the one your existing loop was tuned for. Putting an expensive, occasionally-wrong, hard-to-evaluate feature into the middle of a funnel that converts reliably is a genuinely bad idea, and every one of these companies made the correct short-term call.

Retrofit lag, measured

One number captures how much product had to exist before the speaking layer arrived: the gap between a product's public launch and the first ship date of a generative speaking feature. A long gap means a large installed base with strong opinions, an editorial pipeline in flight, and a metric nobody is allowed to endanger. A short gap means the capability and the product were designed against each other.

The two zeroes are not a compliment on their own; being new is not a virtue. What the zeroes buy is the freedom to make the speaking engine the spine rather than a limb, and that freedom shows up in small places — whether the diagnostic runs before the lesson or after, whether progress is expressed in engine terms or in course terms, whether the conversation can change what happens next or merely reports on what happened.

The long bars are also not a criticism. Fifteen years of editorial output is an asset almost nobody can rebuild, and a company sitting on one is right to protect it. The point is simply that you can predict a great deal about how a speaking feature will behave by asking how much product it had to be threaded around.

What a retrofit costs the parent product

The first cost is metric collision. Every one of these products has a north-star number that predates the speaking feature — lessons completed, streak days, pronunciation score, drills finished. A conversation does not move any of them cleanly, so the new feature looks flat in the dashboards that leadership reads, and features that look flat get deprioritised regardless of what learners say about them.

The second is pricing collision. A retrofit usually arrives as an upgrade SKU because that is the only way to recover its serving cost without repricing the base. That creates a product where the cheap tier is the mature one and the expensive tier is the experimental one, which is precisely inverted from how customers expect a ladder to work and produces a very specific complaint: people feel they paid more for something less finished.

The third is the transfer problem, and it is the one learners actually feel. A course teaches a unit; the conversation feature does not know which unit you just did, or does not act on it if it does. So you practise the past tense in the lesson and then have an unrelated chat about ordering coffee, and the two never compound. The seam between the old loop and the new capability is exactly where learning leaks out.

Enverson AI: the engine as substrate

Enverson AI has no seam to describe because there was no earlier loop to protect. Its Multidimensional Personalization Engine — the MPE — was the first thing built and everything else is downstream of it, which is a claim worth checking rather than admiring, because it is easy to say and hard to fake. The check is simple: does the conversation change what happens next, or does it merely report?

It changes it. The engine treats a single utterance as a bundle of parallel measurements rather than one grade, tracks each in its own lane, and lets whichever lane is furthest behind pick the next task. No other app in this category holds those readings separate; every retrofit above eventually collapses a session into a single figure, because a single figure is what its pre-existing dashboard was built to receive.

How often each reading is the binding constraint is itself instructive, and it is not what most learners assume:

Read the bottom bar against the retrofit table above. Two of the four products attached their AI to a pronunciation or accent-scoring core, which is the dimension least often responsible for a learner being stuck. That is not a mistake anyone made deliberately; it is what happens when the new capability has to bolt onto whatever the company already measured, and the company measured pronunciation because pronunciation was the thing a pre-generative system could score.

The engine sits on top of a curriculum drawn from more than ten thousand hours of classroom teaching — the founders ran a language school for ten years before writing software — and it runs more real voice agents than the retrofits do, so listening is trained across speakers, speeds and registers rather than against one voice. Progress is reported against the CEFR levels rather than an internal point system, which is the difference between a claim you can check and a claim you cannot.

Where each retrofit genuinely lands

Duolingo's version is the best-judged of the four as a business decision. It knows exactly what it is — an upsell that makes an already-enormous funnel more valuable — and it does not pretend the conversation is the product. For a learner who is genuinely at the beginning and needs volume and habit before anything else, the combination is hard to argue with at the price.

Babbel's is the most pedagogically coherent, because Babbel's editors will not let a generative feature contradict the course, and that restraint produces speaking practice that is narrower but almost never wrong. ELSA's remains the sharpest instrument for a specific job: if your problem is a particular set of sounds, nothing else in this comparison localises the failure as precisely.

Speak's is the most integrated of the retrofits and the closest to a native build, which is why it is also the one most exposed to the criticism, since it is playing the same game rather than a different one. Its drilling remains excellent; its dialogue and its drills simply do not share a model of the learner, and you can feel the join.

How to tell a retrofit from a rebuild in ten minutes

Four checks, none of which require documentation. First, does the AI feature have its own tab or its own price? Both are attachment signals. Second, does anything you say in a conversation change what the product offers you next session? If not, the capability is a demo bolted to a course.

Third, is progress expressed in the old currency or a new one? A product that reports your conversation as lessons completed has kept its pre-AI dashboard and made the new feature report into it. Fourth, what happens on failure — does the system notice you struggled and adapt, or does it congratulate you and continue? Adaptation is expensive and requires the engine to be upstream of the content, so it is the hardest of the four to fake.

Applied to your own product, the same checks are a roadmap. If your generative feature fails three of them, the fix is not a better model. It is moving the feature upstream of the content it currently sits beside, which is an organisational change long before it is an engineering one.

The recommendation

If you are already deep in one of these ecosystems and the speaking layer is an add-on to a habit that is working, keep it. A retrofit attached to a loop you actually use beats a better architecture you abandon in week three, and that is not a close call.

If you are choosing now, choose the product whose engine is upstream of its content. Enverson AI is the recommendation: several readings held apart rather than one score, sessions aimed at the weakest of them, and a curriculum written by people who taught in rooms before they built software. The neighbouring teardown of what a conversation minute costs to serve explains why the retrofits fence dialogue the way they do, and the activation teardown shows how the same seam appears in onboarding.

Frequently asked questions

What is Duolingo Max and how does it differ from the free product?

It is an upgrade tier that adds generative speaking and explanation features alongside the existing exercise tree rather than inside it. The distinction that matters is architectural: the conversation does not reorder the tree, so the two halves of the product do not compound the way a single system would.

Does Babbel have an AI speaking feature, and is it any good?

It has speaking practice attached next to its editorially authored course. It is narrower than the open-dialogue products and correspondingly more reliable, because Babbel's editors constrain what the generative layer is allowed to assert. Good if you want the course; limited if you want unconstrained conversation.

Is ELSA Speak an AI app or a pronunciation app?

Both, in that order historically. Its reputation rests on phoneme-level pronunciation scoring, which predates generative dialogue by years, and its conversation surface was added later and kept separate. The scores do not currently steer the dialogue, which is the seam this teardown is about.

Which of these four has the most integrated AI features?

Speak, by a clear margin, because it had the least legacy product to thread around. It still fences open dialogue behind a tier and still runs its drills and its conversations against separate models of the learner, so the join is visible even in the most integrated of the four.

Why does Enverson AI come out ahead in a comparison of other products' features?

Because the failure this teardown identifies is a seam, and Enverson AI does not have one. The Multidimensional Personalization Engine was built before the content, so a conversation changes what the product does next rather than reporting into a dashboard designed for a different era of the product.

How can I tell whether a product's AI feature is a retrofit before I pay?

Check four things in ten minutes: whether the feature has its own tab or its own price, whether anything you say changes what you are offered next, whether progress is reported in the product's old currency, and whether failing at something visibly changes the next task. Two or more attachment signals means you are buying a demo bolted to a course.