The future of language learning: how to build the AI-native Duolingo

There are two ways to effectively learn a new language. You can move to a specific country for complete immersion. Or you can use 1:1 tutoring, which we all know is proven to be the most effective method for learning. Neither comes cheap!

But options that are more affordable and widely accessible also tend to be less effective. You can subscribe to one of the many digital learning players to learn through gamified learning pathways, but they lack speaking practise opportunities, feedback and motivation/accountability, and they feature limited personalisation.

The market is huge. One in four people globally attempt to learn a new language, which means that looking at the total addressable market from the bottom up we come to a figure of $63bn, made up of both B2C ($60bn) and B2B ($3bn).

HolonIQ put the B2C language learning market at $49bn in 2020, growing to $115bn by 2025. So by this point (mid-2024), we should be at ~$105bn, 1.8x our figure in the bottom-up working. For the sake of argument, let’s assume the true market size lies somewhere between the two figures and we arrive at ~$85bn.

According to HolonIQ, only ~40% of this market is currently digital, with the remainder all “offline” (think publishing players such as Pearson, and live learning players such as Education First).

This represents a huge opportunity for digital disruption, especially given the capabilities of AI.

Read on for:

• An overview of the language learning market, including a market map and anatomy of existing incumbents and their business models
• Key insights into the top 10 “first movers” in AI-powered language learning
• Specific ideas about what to build and advice on the differentiators we look out for as investors in this space.

* * * 

AI is powering the following new frontiers of innovation for language learning…

Real-time conversational practice

Problem: having someone to talk/practise with, who knows who you are (name, interests, etc) and your ability level, is a great driver of conversational/spoken progress.

Solution: AI tutors unlock new levels of conversational practise for students, who can engage in real-time conversations with their AI tutors in foreign languages, perfecting their speaking skills.

Who’s doing this?

• Speak – Offers AI powered conversational practice tool
• Google – Offers AI powered conversational practice toolal practice tool for language learners
• Open AI
  • GPT-4: there is no UI enabling language learners to learn verbally via an AI tutor. The free talk function essentially just converted speech to text, interfacing with the LLM and then converting text to speech for the response. This meant there was no ability to provide critique / feedback on how you speak.
  • GPT-4o: with the release of GPT-4o, the model takes your speech and ingests it directly, opening up new possibilities for language learning applications (predominantly feedback on pronunciation).

What does this mean for startups?

• Our take is that this is a great opportunity for startups, as Open AI cannot feasibly create application layers across all of the use cases for the models they are building. GPT-4o unlocks new possibilities for startups to leverage the work of Open AI to ship great products for learners, fast.
• See the language learning example GPT-4o videos on point-and-learn Spanish and real-time translation

Hyper-personalisation

Problem: lack of personalisation from existing digital language learning solutions. For example: relatively generic learning pathways, irrespective of specific development areas or interests.

Solution: AI can unlock new depths of personalisation. For example: hyper-personalised learning pathways; continuously mapping user level, interests and learned words/phrases to content; using vocabulary and grammar you recently learned in subsequent exercises to continue momentum.

Who’s doing this?

Praktika – Offers hyper-personalised onboarding pathway with accurate ability level placing.

Having selected that my primary reason for learning English is for career growth, Praktika then asks specifically what I am looking to leverage English for in my career.

The app then asks me to specify my hobbies and interests to tailor specific exercises and content to my preferences.

Ability to choose the accent I am looking to speak English with.

All of these preferences then feed into a hyper-personalised learning pathway.

Pronunciation

Problem: conversational practise is great but solves only part of the problem. As users advance, they will not only want to be able to construct sentences but work on specific pronunciation challenges. Traditionally, cracking pronunciation requires personalised feedback and specific exercises, ordinarily from a human tutor.

Solution: AI-powered pronunciation practice, tailored to your specific background and needs.

Who’s doing this?

BoldVoice – Offers AI powered pronunciation

The app asked me for my native language and then personalised exercises based on known pronunciation difficulties for that specific language. It offered highly specific pronunciation exercises, with examples of good and bad pronunciation.

Speaking exercises using examples.

Feedback on performance, ability to listen to your response vs best in class response side by side.

Taking real-world content and levelling it to the user

Problem: content is generic and time consuming to produce so there is a lot of content recycling. But what is engaging and relevant for one user is likely not for another.

Solution: using AI to take real-world content (news, movies, tv shows), categorise the clips based on words used and “difficulty” (speed of speech), and then deploy it as personalised learning content for the user. In addition to being engaging for the user, this content can also be culturally immersive without being too difficult. Cultural immersion provides the double benefit of helping the user learn both the language and about local culture.

Who’s doing this?

Wordbox – Uses real-world content (tv, movies, news) and tailors it to the language learner’s level.

Wordbox ingests real-world content, categorises it and then uses it for language learning, selecting specific content based on the language learner’s ability and learning stage.

Accountability

Problem: given the lack of human interaction, app-based learning can be unengaging in comparison to 1:1 tutoring and leaves the learner with a lack of accountability and motivation. While Duolingo and others have mastered retention via game-driven mechanics of streaks, it is arguably a lower pull than missing a live lesson with your classmates / teacher (which in many cases you have pre-paid for).

Solution: AI tutor-led language learning combines (in theory) the best of both worlds – accountability to a tutor and game-driven mechanics. The question of whether we will feel accountable to AI tutors is certainly still an unanswered one and we are interested to see how this plays out.

Who’s doing this? 

From our research, only Speak has launched something near to this with its AI tutor, which can generate exercises and provide feedback live. This is the closest to “accountability” we’ve found, but is still in its very early stages and there is a lot more to do here!

Market landscape

Having explored how AI is creating opportunities in this space, let’s now look at today’s market landscape and assess where these AI-powered frontiers of innovation can disrupt market segments and create opportunities for startups.

Emerge has deeply researched the market to create this unique map of the language learning market, detailing 135 language learning companies split by scalability, interactivity and user segment. The map moves from unscalable and static at the bottom left through to scalable and live at the top right.

Let’s look at each segment in turn.

Segment A (Static X Unscalable)

• n/a: Nature of static content is that it is scalable, hence no companies are highlighted in this space.

Segment B (Static X Scalable)

Summary: This segment covers language content and assessment players. These are typically high margin businesses dominated by key players such as IELTS, PTE and TOEFL. However, in more recent history we have seen market entry from Duolingo with its “Duolingo English Test”, which has been widely accepted and is a signal that there is opportunity for new players here.

Pros:

• Assessment/certification form a crucial part of the language learning process.
• For a large portion of the market, it will be in some way required: for kids (getting into schools), students (getting into universities) and professionals (getting jobs), being able to prove that you have the skills is fundamental.
• Duolingo succeeded in creating new standard with the DET, which is an exciting precedent for future language unicorns.

Cons:

• It is challenging for companies to gain the recognition needed to become an accepted form of language proficiency.

Where is the exciting opportunity?

Segment C (Gamified X Unscalable)

• n/a: Nature of gamified content is that it is scalable, hence no companies highlighted in this space.

Segment D (Gamified X Scalable)

Summary: This segment covers gamified language learning apps. These companies pioneered mass market affordable language learning, capitalising on the transition to mobile. These players are typified by gamified content with structured learning pathways. However, they generally lack the level of personalisation and opportunities for conversational practice required to take learners to the next level of ability.

“There is a big opportunity for disruption in the B2B space. Casuals has had a lot of competition, but legacy incumbents still dominate the professionals market.”

 

Gustaf Nordback, ex-MD EMEA & Brazil, Rosetta Stone, and Emerge Venture Partner

Pros:

• Low cost and democratising access to language learning.
• Gamified experience is highly engaging, which can lead to habit forming and higher retention.
• Apps don’t tire. Highly motivated students can rack up significant hours learning languages v. limited point-in-time in-person lessons.

Cons:

• Generally, the aim of gamified apps is not fluency, but getting beginners to limited proficiency. Learners reach a ceiling.
• Limited speaking practise.
• Limited levels of personalisation.
• Limited accountability due to lack of human tutor.

Where is the exciting opportunity?

Segment E (Live X Unscalable):

“Will AI replace trainers? No, but trainers with AI will replace trainers without AI.”

Benjamin Joseph, cofounder & CEO Learnlight, and Emerge VP

Summary: this segment covers marketplaces and tutoring platforms. They pioneered the early models of language learning, offering generally higher cost education for those with the ability/propensity to pay for human centric learning

Pros:

• Highly personalised learning, with tutors able to tailor content to learner needs on lesson-by-lesson basis
• Speaking practise and cultural immersion (through conversing with locals and cultural language travel)
• In general, excellent pedagogy (but can be highly tutor dependent, so not guaranteed)
• Highly engaging, with human element resulting in motivation and accountability for learners
• Certification via language school, which offers credibility

Cons:

• Cost is high and therefore lower access for the majority of individuals
• Lack the “always on” element that gamified app-based players have, meaning less attractive to casual learners
• It is daunting speaking in front of a human, which can be offputting for some learners

Where is the exciting opportunity?

Segment F (Live X Scalable)

“AI will remove the ceiling traditionally faced by more advanced language learners with regards to app-based learning.”

 

Bernhard Niesner, cofounder & ex-CEO Busuu and Emerge VP

Summary: this segment represents the future of language learning and where we are looking to back the next unicorn. Live learning powered by AI, available at a fraction of the cost of traditional live learning solutions, unlocks access to all the benefits of 1:1 tutoring at a price point in reach for the mass market. The opportunity here is huge, and there is the chance to capture value from almost all the existing segments of the language learning market.

Pros

• Real-time conversational practice: chat with AI through role play or with an AI tutor
• Hyper-personalisation: everything personalised to you. If the goal is a work presentation, then have your learning pathway tailored accordingly
• Pronunciation: AI-powered feedback on how you speak, to help you sound like a native
• Taking real-world content and levelling it to the user by leveraging content the user loves
• Accountability: feel a drive to turn up and pick up where you left off with your AI tutor

Cons

• Novelty wears off: learners might get tired of interacting with an AI over time, potentially leading to decreased engagement and reverting to other methods
• High vulnerability to app-based incumbents: new players must replicate the gamified learning pathways of established apps, such as Duolingo, while also building AI live learning features, making them highly vulnerable to established app-based incumbents
• Cost of AI features: being AI-first can come with increased costs. Companies need robust justifications for these costs and clear evidence of how AI improves learner outcomes
• Market dynamics: casual users might drop out of the market due to advances in real-time translation technology, reducing the perceived need for language learning

Given this segment is too new to have incumbents, below you can see Emerge’s overview of the top funded “AI-first” language learning apps.

Where is the exciting opportunity?

Conclusions

While startups have significant opportunities in this category, it is not without its challenges. We see the top three challenges as:

1. Incumbent threat

• Challenge: established companies can quickly adopt AI features (either in-house or with a small acquisition), posing a significant threat to new entrants. Ultimately it will come down to who is faster: incumbents launching AI features or startups achieving enough scale off the back of their headstart with AI.
• Emerge view: early movers in the AI language learning space are scaling hyperfast at the moment and our view is that they will be too big to fail in two to three years. However, unless you are building immediately and investing heavily into GTM, it will be too late.

2. Cost of AI features

• Challenge: being AI first can come with increased cost, so there will be product and user experience trade offs.
• Emerge view: we believe that to win you need to be AI-first at the beginning to gain market share, and over time scale back and become more “Duolingo” in the future with different paid tiers.

3. Market positioning

• Challenge: Duolingo attracts casual language learners. There is a question mark over whether these learners care about spending time on AI-powered features to take them to the next level, such as conversational speaking practise, pronunciation feedback, etc. Casual learners may start caring about these features because they’re now cheap enough that it no longer represents a serious financial commitment. However, in another world, the AI features will be used by serious language learners only – and they represent a much smaller segment.
• Emerge view: we believe that AI lowers the cost of serious learning to the point where casual learners will interact with the features so long as they are fun and engaging