ELSA, Corp

Here is a fact about language that most language apps would prefer you not think about too hard: you can know a language perfectly and still be unable to use it. You can have the vocabulary, the grammar, the reading comprehension of a native speaker, and then get on a phone call and be asked, gently, to repeat yourself. Again. This is a specific and quiet form of failure, and it does not show up on any test. ELSA, Corp is a company built entirely around it.

The name is an acronym - English Language Speech Assistant - which tells you the founders were engineers before they were marketers. What ELSA actually does is narrower and stranger than "language learning." It listens to you speak English and then tells you, phoneme by phoneme, where your mouth went wrong. Not your spelling. Not your word choice. The physical sounds. It is the difference between a teacher who reads your essay and a coach who watches your feet.

The origin story is the kind that sounds too neat until you realize it is also the entire product thesis. Vu Van, ELSA's co-founder and CEO, left Vietnam for Stanford, collected an MBA and a Master's in Education, and by every credential was fluent in English. She was also, by her own account, constantly misunderstood in conversation. The gap between "passed the test" and "got understood on the call" was the gap she decided to close. To close it she needed someone who understood the physics of speech, so she recruited Xavier Anguera, a speech-recognition scientist. A businessperson who lived the problem and a scientist who could hear it. That is the whole company.

The Trick Is the Training Data

If you want to understand why ELSA works when other pronunciation tools feel like toys, you have to understand one deliberately weird design decision. Most speech-recognition systems are trained on native speakers, because native speakers are the "correct" data. ELSA did the opposite on purpose. It trained its models on the voices of non-native speakers - people carrying Vietnamese, Korean, Portuguese, and Spanish accents into English. The result is a system that has heard, thousands of times, the exact way a particular accent bends a particular sound, and can therefore tell you what to move.

This matters because a native speaker's ear is, counterintuitively, a bad coach. A native listener's brain does enormous unconscious repair work - it smooths over your errors, fills in the sounds you missed, and hands you a "good enough" that leaves you no wiser. ELSA refuses to do the repair work. It has been described as detecting pronunciation errors with better than 95% accuracy, and the honesty is the feature. The app highlights the failed syllable in red and gives you a score you did not ask for. There is no confetti. The bet - and it is a genuinely interesting bet - is that adults do not want to be flattered by a coach. They want to be told what to fix.

One Engine, Four Businesses

The consumer app is what people know, and it grew the way good consumer apps grow - by word of mouth, in the tens of millions, across more than a hundred countries, on the back of a freemium subscription called ELSA Pro. The company cites figures in the range of 50 million downloads and 1.6 million reviews. But the more instructive part of the ELSA story is what it did with the speech engine once it existed, because building an AI that can hear an accent is expensive and hard, and once you have paid for it you would be foolish to sell it only once.

So ELSA sells it four ways. There is the consumer app. There is ELSA Business, which takes the same engine into companies as workforce-enablement software - useful anywhere employees need to be understood on international calls. There is ELSA Schools, packaged for educators with analytics and teacher resources. And there is the ELSA API and Speech Analyzer, which lets other organizations embed the pronunciation-scoring technology directly. The hard part, the AI that hears accents, only had to be built once. Everything after that is leverage.

Then generative AI arrived, and ELSA had an advantage most edtech companies did not. When the industry rushed to bolt chatbots onto everything, ELSA already owned something proprietary: it knew how you sound. In September 2023 it launched ELSA AI Tutor, a generative-AI conversation partner that does not merely chat but holds an open-ended spoken conversation and corrects your pronunciation in real time. That is a materially better product than a text chatbot, and it exists because ELSA had spent years accumulating the one kind of data a language chatbot cannot fake - the sound of your actual voice.

The Money and the Map

ELSA has raised roughly $60 million across four rounds, and the cap table is worth reading because it crosses three continents. An early seed round drew Monk's Hill Ventures. The 2019 Series A was led by Gradient Ventures, Google's AI-focused fund - a useful stamp for a company whose entire claim is "our AI is good." A Series B followed. Then in September 2023 came the $23 million Series C, led by UOB Venture Management, with participation from UniPresident, Japan's Aozora Bank and Development Bank of Japan, and Vietnam Investments Group, alongside returning backers.

You do not usually see a Singapore bank's venture arm, a Japanese bank, Google's AI fund, and a Vietnamese growth investor on the same line item. But it fits the shape of the problem. English learning is not a Silicon Valley market - it is a global one, concentrated in Asia and Latin America, and ELSA's investors map onto exactly the geographies where being understood in English is a career decision rather than a hobby. Third-party estimates put ELSA's revenue around $32 million with a team of roughly 200 to 250 people, largely split between San Francisco and Vietnam.

The Competitor Everyone Names, and the Fight ELSA Didn't Pick

Every comparison of ELSA eventually invokes Duolingo, and the comparison is instructive precisely because the two companies are not really doing the same thing. Duolingo is a gamified, level-based system that teaches vocabulary and grammar across many languages, and it is very good at getting you to show up. ELSA does one language and starts from an assumption Duolingo does not make: that you already know English. It is built for the intermediate-to-advanced learner who can read the newspaper but cannot get through a meeting without being asked to repeat a word. ELSA did not try to beat the category leader. It took the part the category leader skipped - the last mile of actually being understood - and made that the whole company.

Whether that is a big enough business to justify the ambition is the open question, and it is a fair one. Pronunciation is a narrower wedge than "learn a language." But narrow wedges into enormous markets have a way of being underestimated, and the market here is not small: something on the order of 1.5 billion people are learning English right now, and vanishingly few of them are doing it to pass a test. They are doing it to be understood at work.

A Company Built By People Who Live the Problem

There is a structural reason ELSA's product feels lived-in, and it is worth stating plainly: the people building it are, in large part, the people it was built for. Leadership and engineering are split between San Francisco and Vietnam, which means the team that trains a model to understand a Vietnamese accent in English is frequently composed of people who speak with one. That is not a diversity footnote - it is a feedback loop. When your engineers are also your hardest test cases, the distance between "we shipped a feature" and "the feature actually helps someone be understood" collapses to about the length of a hallway. Vu Van has built the company around that lived frustration rather than around a market-sizing spreadsheet, and it shows in the choices: the refusal to flatter, the focus on the specific sounds a specific accent gets wrong, the decision to treat honesty as the headline feature rather than a bug to be smoothed over.

The leadership bench has grown up around that thesis. In 2025 the company named Akshaya Aradhya - previously an AI leader at GitHub, Netflix, and LiveRamp - as chief technology officer, a hire that reads as ELSA taking its "our AI is genuinely good" claim seriously enough to staff it at the top. Around the executive table are veterans from Udemy, Google, and Uber, which is the sort of resume list that a consumer-plus-enterprise company assembles when it intends to sell to individuals and institutions at the same time. None of it changes the fundamental bet, which remains almost stubbornly simple.

ELSA has wagered its entire existence on the idea that intelligibility is a measurable, coachable skill - and that people will pay to be told the truth about how they sound. Recognition has followed the bet, including a World Economic Forum Technology Pioneer designation in 2024 and a place among the AI companies the business press likes to list. The harder validation is quieter and harder to chart: whether, after a few weeks, the red syllables start turning green. That is the one number ELSA cannot buy with a funding round, and it is the only one that ultimately matters.