ELSA Speak

Every non-native English speaker knows the specific humiliation of the repeat. You say a word. The other person tilts their head. You say it again, louder, as if volume were the problem. It never was. The problem is a phoneme — one small, specific, learnable sound — and no one in your life has ever been rude enough to tell you which one.

That’s the gap ELSA Speak was built to fill. Developed by ELSA Speak (the name is a backronym: English Language Speech Assistant), the app is an AI-powered pronunciation and speaking coach aimed squarely at learners who already have vocabulary and grammar but can’t get their mouth to cooperate. Its origin story is the interesting part: the company was co-founded by a Vietnamese engineer, Vu Van, alongside speech-recognition researcher Xavier Anguera, and the founding insight shows up in the product architecture. Most speech recognition is trained on native speakers, which is why it hears an accented learner and either guesses charitably or fails outright. ELSA’s recognizer is trained on the speech of language learners — it’s built to hear the error rather than smooth over it.

Market-wise, ELSA sits in an awkward and rather valuable spot. Duolingo owns the top of the funnel and treats speaking as a garnish. Babbel and Busuu do structured curricula with human-ish conversation. Italki sells you an actual person. ELSA ignores all of that and specializes: it is a pronunciation instrument first, with a language curriculum stapled around it. 8,000+ bite-sized lessons, multi-accent support (American and British), scenario practice for job interviews and meetings, and exam pathways for IELTS, TOEFL, TOEIC, and CEFR.

The reception backs the thesis. Roughly 4.6 on the App Store and 4.5 on Google Play — genuinely strong for a category where users are constantly disappointed. The praise clusters tightly around the feedback engine. The complaints cluster just as tightly around the paywall and the occasional moment when the recognizer is confidently, maddeningly wrong. Both of those are real, and both matter.

• Phoneme-Level Feedback: The core engine. Speak a sentence, get it back color-coded by sound — not by word, by sound. It isolates the specific phoneme you flubbed and shows you how to fix it. This is the product.
• Stress, Intonation & Fluency Analysis: Beyond individual sounds, it scores syllable stress and sentence-level intonation — the two things that make technically-correct English still sound wrong.
• 8,000+ Bite-Sized Lessons: Structured pathways across pronunciation, vocabulary, grammar, and listening, tuned for short sessions rather than marathon study blocks.
• Multi-Accent Targeting: Choose American or British English as your target model instead of being forced into a single default.
• Scenario & Exam Prep: Job interviews, meetings, presentations, plus dedicated tracks for IELTS, TOEFL, TOEIC, and CEFR-aligned assessment.
• Overall Speaking Score: A single tracked number that trends over time, with adaptive difficulty adjusting the daily set to your level.

The Feedback Loop

The loop is dead simple and that’s the point. Prompt appears. You tap the mic. You speak. Within a beat, the sentence comes back to you highlighted — green where you nailed it, yellow where you were close, red where you weren’t. Tap the red part and you get the diagnosis: which phoneme, what it should sound like, where your tongue is supposed to be, and a native reference to imitate.

The latency matters more than the feature list suggests. Pronunciation is a motor skill, not a knowledge problem. You are training muscles, and muscle training requires feedback fast enough to connect to what your mouth just did. A human tutor gives you this feedback once a week for money. ELSA gives it to you forty times in ten minutes for free-ish, at 11pm, with no one watching. The absence of social cost is the actual innovation here — learners will attempt sounds in front of an app that they’d never risk in front of a person, and repetition is where pronunciation is won.

The color-coding is also a smarter interaction design choice than it appears. It converts an abstract, invisible failure (“your accent”) into a discrete, tappable object (“this /θ/”). That reframing is most of the value. Suddenly you don’t have a bad accent — you have eleven specific sounds to fix, and a list is something a person can act on.

Where the Machine Breaks

Now the honest part. The recognizer is not infallible, and users say so consistently. It will occasionally mark a correct pronunciation as wrong, or wave through something you know was sloppy. This is not a bug so much as the fundamental condition of the category — the model is scoring a probability distribution over acoustics, not reading your mind — but the product handles those moments poorly. When the app is wrong, it’s wrong with the same crisp confidence it has when it’s right. There’s no visible uncertainty, no “I’m not sure, try again.” For a learner without the expertise to overrule the machine, a false negative doesn’t just annoy — it can teach them to un-learn something they already had correct. An app whose entire authority rests on precision should be more forthcoming about the edges of its precision.

Mic quality and background noise compound this. Practice on a noisy commute and the scores get noticeably less trustworthy, which quietly undercuts the app’s own pitch about short sessions squeezed into dead time.

The Curriculum Problem

The engine is a scalpel. The curriculum around it is a blunter tool. Early on, the adaptive difficulty feels genuinely responsive — it finds your weak phonemes and drills them. But push into the higher levels and the seams show: reviewers repeatedly flag repetitive content, the same structures recycled with different nouns. 8,000 lessons is an impressive number and a slightly misleading one, because past a certain point you aren’t learning new material, you’re doing reps. Which is arguably correct pedagogy for pronunciation! But it’s not what the marketing number implies, and the app doesn’t frame it honestly enough for advanced users to know they’ve graduated from learning to maintenance.

The exam prep is the strongest part of the surrounding curriculum, and not by accident — IELTS and TOEFL speaking sections are scored on exactly the dimensions ELSA measures. The alignment between what the app is good at and what the test rewards is unusually tight. If you’re paying for ELSA for one reason, this is the most defensible one.

The Paywall

The free tier gives you a handful of daily lessons and then a wall. That wall is the single most common complaint in the reviews, and it’s aggressive by design: you’re gated right at the moment the feedback engine has hooked you. Premium unlocks the full curriculum, advanced assessment, and exam courses. It is not cheap, and the pitch you’re being asked to accept is that it’s cheaper than a tutor — which is true, and also the argument every subscription app makes right before it starts feeling like a utility bill. The counterargument in ELSA’s favor is unusually strong, though: unlike most subscription language apps, the thing behind the paywall is not content, it’s a measurement instrument you can’t get anywhere else at that price.

The design is competent and unremarkable, which here is close to the right call. Clean cards, generous tap targets, legible type, a progress dashboard that leans hard on the big central speaking score. It doesn’t chase the aggressive gamified-dopamine aesthetic that Duolingo made the genre default, and the restraint reads as maturity — this app thinks you’re an adult with a job interview coming up, not a child who needs a streak to survive.

The one place the aesthetic genuinely earns its keep is the feedback screen. The phoneme highlighting is clear, well-spaced, and readable at a glance, which is exactly what a screen you’ll look at several hundred times needs to be.

Performance is where the real engineering lives, and it’s mostly invisible — which is the compliment. Speech capture and scoring return fast enough to feel conversational. Sessions are short by design, so the app is rarely on screen long enough to strain anything. Battery and thermals are a non-issue for the intended 10-minute use case.

Both iOS and Android deliver the same core product, which is the right decision and, given ELSA’s user base, a commercially necessary one — a large share of its learners are in markets where Android is not the secondary platform, it’s the only one.

The meaningful difference isn’t the OS, it’s the hardware underneath it. Speech scoring is only as good as the microphone feeding it, and the iOS fleet is narrow and consistent while the Android fleet spans flagship glass down to hardware where the mic is the last thing anyone thought about. That variance lands directly on the feature the whole app rests on. Practical advice, platform-agnostic: wear headphones with an inline mic, and treat a sudden unexplained dip in your score as a hardware or noise problem before you believe it about your mouth.

Neither version is a lazy port. Both respect their platform’s navigation conventions well enough to disappear.

ELSA Speak is a specialist tool that knows precisely what it is, and I’d rather have that than another app trying to be everything. The pronunciation engine is the real thing — not a demo, not a checkbox, but a genuinely useful instrument that hands learners something they’ve never had access to at this price: an honest, immediate, specific answer to “what am I actually doing wrong?”

It’s held back by two things it could fix and hasn’t. The recognizer’s false confidence deserves a humility layer — a simple “low confidence, try again” would cost the company nothing and save learners from occasionally being taught the wrong lesson by a machine. And the paywall is tuned for conversion rather than goodwill, clamping down at exactly the moment a learner starts believing this might work.

Neither flaw is fatal, because the core is that strong. If you can name the thing you want to fix about your spoken English, ELSA can probably hear it, and that’s more than your colleagues will ever tell you.

7.9/10 — a sharp, focused tool with a monetization strategy that doesn’t quite deserve it.

One note outside the review, since it may matter for the pipeline: everything above is grounded in the supplied description and research notes, with the exception of the founder background in the Overview (Vu Van and Xavier Anguera), which I drew from general knowledge rather than the input data. It’s well-documented publicly, but if this pipeline requires all claims to trace back to supplied research, that sentence is the one to cut or verify.