Quick summary: Speak is the AI-powered language learning app founded in 2016 by Connor Zwick and Andrew Hsu (both Thiel Fellows). Originally launched targeting English learners in South Korea, it expanded globally and now teaches six languages: Spanish, French, Korean, Japanese, Italian, and English. Speak is conversation-first — the central practice is speaking out loud to an AI that responds with corrections on tone, pronunciation, and grammar. The OpenAI Startup Fund is an investor; Speak uses OpenAI’s Realtime API for the conversation engine. 15 million+ downloads, 4.8 App Store rating, $1B valuation as of late 2024. Pricing: $20/month or $99/year. The strongest AI-only conversation-practice product on the market — meaningfully better at conversation than Duolingo Max. Updated 2026-05-16.
If you want to actually become conversational in a new language using AI, Speak is the product most language teachers will quietly recommend. It is not the most-downloaded language app (Duolingo wins that race by an order of magnitude). It does not have the most languages (most catalog apps win on quantity). It is, however, the app most committed to a single thesis — that the bottleneck in adult language learning is not vocabulary or grammar but the act of speaking out loud — and the app that has built the conversation engine most committed to that thesis. This post is the practical guide for anyone choosing among AI conversation tools, and the honest comparison with Duolingo Max, ChatGPT Voice, italki, and the rest of the stack.
Get Smarter About AI Every Morning
Free daily newsletter — one story, one tool, one tip. Plain English, no jargon.
Free forever. Unsubscribe anytime.
What is Speak in plain language?
Speak is a mobile app (iOS and Android) that teaches languages through structured AI-powered speaking practice. The central loop: open the app, pick a lesson, the app prompts you in the target language with a phrase or question, you say something back out loud, the AI evaluates your pronunciation, fluency, and grammar, and the conversation continues. Where Duolingo and Babbel give you vocabulary drills and Khan Academy Kids gives you reading practice, Speak gives you the actual speaking practice that the adult language learner most needs and most lacks.
The product has four main modes:
- Video Tutor lessons. Short structured lessons taught by a video instructor with embedded speaking practice. The lessons build progressively from beginner through advanced and cover core grammar and vocabulary alongside the speaking work.
- AI Tutor conversations. Open-ended conversation with an AI tutor on any topic. You set a topic; the AI talks with you in the target language, corrects when needed, and adapts to your level.
- Live Roleplays. Scripted real-world scenarios — ordering coffee, job interview practice, asking for directions, negotiating a price — with AI characters. After each roleplay you get specific feedback on accuracy and use of varied language.
- Speak Tutor for specific topics. Targeted practice on grammar points, vocabulary themes, or business-specific scenarios.
The technical layer underneath: Speak uses OpenAI’s Realtime API (the same low-latency voice infrastructure behind ChatGPT Voice) for the conversation engine, combined with Speak’s own curriculum and feedback systems trained on language-learning data. The result is the lowest-latency, most natural-feeling AI conversation in the language-learning category as of 2026.
Who founded Speak and what’s the story?
Speak was founded in 2016 by Connor Zwick and Andrew Hsu, both Thiel Fellows. The Thiel Fellowship is Peter Thiel’s $100,000 grant program for young entrepreneurs who skip college to build companies. Zwick and Hsu were classmates in the same Thiel Fellowship cohort. The original product was an English-learning app targeting South Korean students — a market Zwick had identified through earlier research as both large and underserved by existing apps.
Speak grew steadily through 2016-2022, dominated the Korean English-learning market, and began international expansion in 2022-2023 with Spanish, French, Italian, and Japanese added to the catalog. The 2023-2024 wave of generative AI was a major inflection point — Speak was one of the earliest OpenAI Startup Fund portfolio companies and got priority access to GPT-4 and the Realtime API. The app’s AI Tutor and Live Roleplay features were rebuilt on this infrastructure and launched as differentiators against Duolingo Max and other competitors.
Total funding raised: ~$162 million across seven rounds. Valuation: $1 billion as of December 2024 ($78M Series C). Investors: OpenAI Startup Fund, Accel, Khosla Ventures, Y Combinator. The company is profitable in unit economics, with average user spending of $99-$240/year and retention strong enough to sustain the valuation.
Which languages does Speak support?
- Spanish (from English)
- French (from English)
- Italian (from English)
- Japanese (from English)
- Korean (from English)
- English (from Korean, Spanish, Japanese, and several other languages — the original product line)
If you’re learning a language Speak doesn’t support — Mandarin, German, Portuguese, Arabic, Russian, Hindi, Vietnamese, Thai, Polish, Dutch, etc. — you’ll need another tool. ChatGPT Voice ($20/month) supports essentially every major language and is the natural fallback. Duolingo covers 40+ languages but with the limits we covered in our Duolingo Max post.
What does Speak cost?
| Plan | Cost (2026) | Notes |
|---|---|---|
| Free trial | Typically 7 days | Full premium access during trial |
| Monthly subscription | $20/month | Full access to all features |
| Annual subscription | $99/year ($8.25/mo equivalent) | Substantial discount over monthly |
| Speak Premier (higher tier) | ~$240/year | Adds additional features depending on the market and current product packaging |
The annual subscription at $99/year is the right price point for serious learners. It’s roughly $25 less than the annual price of Duolingo Max and meaningfully better for actual conversation practice. Compared to a human italki tutor at $10-$30/hour (which would cost $250-$750/year for one weekly session), Speak is dramatically cheaper for daily practice volume.
How does Speak compare to Duolingo Max, ChatGPT Voice, and italki?
| Tool | Cost | Best for | Conversation quality | Best language |
|---|---|---|---|---|
| Speak | $99/yr | Daily AI conversation practice with structured feedback | Excellent; purpose-built; OpenAI Realtime API | Spanish, French, Korean, Japanese, Italian, English |
| Duolingo Max | $168/yr | Duolingo users who want a conversation layer integrated with the existing curriculum | Good; curriculum-aware | 7 languages (Max-supported) |
| ChatGPT Voice (Plus) | $240/yr | Flexible conversation in any language; uses the same Realtime API | Excellent; broadest language coverage | Essentially any major language |
| italki (human tutors) | $10-$30/hour | The irreplaceable human conversation partner | Real human; depends on tutor | ~150 languages |
| Pimsleur | ~$165/yr | Listening-focused commute learning; audio-only | Pre-recorded; not interactive AI | ~50 languages |
| Tandem / HelloTalk | Free | Community language exchange with native speakers | Real humans; quality varies | Most languages |
The honest synthesis for serious learners: Speak is the best AI-only conversation tool for the languages it covers. For languages outside its catalog, ChatGPT Voice is the right substitute. For genuine speaking improvement, layer AI conversation (Speak or ChatGPT Voice) with at least some human conversation (italki, Tandem, or in-person). The combination of daily AI practice plus weekly human conversation is the configuration that produces measurable conversational fluency in 6-12 months for committed adult learners.
What does Speak do better than its competitors?
- Conversation latency. Speak’s use of OpenAI’s Realtime API produces near-natural conversation pacing. Pauses feel right. The AI doesn’t take 4 seconds to respond. This sounds minor and is the single biggest determinant of whether the conversation feels useful vs frustrating.
- Pronunciation feedback specificity. Speak’s models give specific feedback on which sounds you got wrong, not just “your pronunciation was 73%.” A learner of Korean gets correction on aspirated vs unaspirated consonants. A learner of Japanese gets feedback on pitch accent. This is purpose-built and shows.
- Roleplay scenario design. The Live Roleplays are well-written real-world contexts — ordering coffee, hailing a taxi, asking for directions, talking to a doctor, ending a phone call politely. Each has specific vocabulary targets and grammatical structures the learner practices in context.
- Curriculum integration. The video lessons and AI conversations are tied to a coherent curriculum that builds skill over weeks and months. ChatGPT Voice has no curriculum.
- Korean and Japanese specifically. Speak’s origin in the Korean market means its Korean and Japanese implementations are particularly strong — better than the equivalent depth in most competing apps.
Where does Speak fall short?
- Limited languages. Six languages is dramatically fewer than Duolingo’s 40+. If you’re learning Mandarin, German, Russian, Arabic, Hindi, or most other languages, Speak is not an option.
- Not the strongest for absolute beginners. The conversation-first approach assumes some baseline vocabulary. Total beginners benefit more from pairing Speak with a vocabulary-building tool (Anki, Duolingo’s free tier, a textbook) for the first few weeks.
- Reading and writing get less attention. Speak’s focus is speaking and listening. Reading-heavy and writing-heavy curricula (Japanese kanji study, Chinese characters, Arabic script) need supplementary tools.
- Real culture and accent variation are limited. AI conversations represent a kind of standardized speaking-language register. Real-world conversations include regional accents, slang, generational vocabulary, and cultural context that an AI tutor smooths out.
- Subscription continuity matters. The compounding value comes from daily use over months. Inconsistent users underperform their subscription value.
- Children-targeted use is limited. The interface and content assume an older teen or adult learner. Younger children’s language learning is better served by other tools.
How should you actually use Speak?
- Start with the free trial. 7 days is enough to feel whether the daily speaking practice fits your learning style and confidence level.
- Commit to daily 15-20 minute sessions. Speak rewards consistency. Three 20-minute sessions a week underperforms five 12-minute daily sessions.
- Use Video Tutor lessons + AI Tutor conversations together. The video lessons introduce vocabulary and structures; the AI Tutor conversations let you practice them immediately. The combination is the highest-leverage use of the product.
- Don’t avoid the Roleplay scenarios. The discomfort of practicing a job interview in your target language is exactly the practice you need before doing it in real life.
- Pair with handwritten study. Per our handwriting research post, take handwritten notes during or after sessions on new vocabulary and grammar. The screen practice is one piece of the puzzle.
- Layer in human practice eventually. An italki tutor for one weekly hour, a language exchange partner via Tandem or HelloTalk, or in-person language meetups. The human practice does what AI can’t and reveals where you actually stand.
- Track progress externally. Speak’s internal progress tracking is fine. For real progress markers, take a CEFR-aligned proficiency assessment (Cambridge English, DELF for French, DELE for Spanish, JLPT for Japanese, TOPIK for Korean) every 6-12 months.
How does Speak fit homeschool foreign-language curricula?
Well, for older students (typically 12+). The conversation-first format requires a baseline reading and listening capability the younger language student usually hasn’t built yet. For homeschool middle and high school students learning Spanish, French, Italian, Korean, or Japanese, Speak is one of the strongest single tools for the conversation component of a language curriculum.
A typical homeschool stack for high-school foreign language: a real textbook (Assimil, Pimsleur, college-level textbook), Anki or Memrise for vocabulary, Speak for daily conversation, real reading material at the student’s level (graded readers, news articles, eventually books), and ideally an italki tutor or co-op conversation partner for the human element. Speak handles the conversation-practice piece better than other AI tools in its supported languages.
For Latin, German, Mandarin, Arabic, Russian, or other languages outside Speak’s catalog, substitute ChatGPT Voice for the conversation-practice piece. The rest of the stack stays the same.
Learn Our Proven AI Frameworks
Beginners in AI created 6 branded frameworks to help you master AI: STACK for prompting, BUILD for business, ADAPT for learning, THINK for decisions, CRAFT for content, and CRON for automation.
Frequently asked questions
Is Speak worth $99/year?
For a committed daily user, easily — about 25 cents per session. Compared to a human tutor at $10-$30/hour, Speak buys you orders of magnitude more practice volume. For sporadic users, the value drops fast. The 7-day free trial tells you which category you’re in.
Speak or Duolingo Max for serious learning?
Speak. The conversation quality is meaningfully better, the curriculum is more conversation-focused, and the price is lower. Duolingo Max is the right choice for someone already deep in the Duolingo ecosystem who wants conversation as an add-on; Speak is the right choice for someone serious about speaking the language.
Can I learn a language with Speak alone?
To genuine conversational fluency — no, not alone, for the same reason no single app gets you there alone. To meaningful progress and basic conversational ability — yes, with consistent daily use over 6-12 months. The full path requires a textbook or structured course for grammar depth, vocabulary work (Anki, Memrise, or graded reading), listening exposure (podcasts, films, music), and human conversation. Speak handles the conversation-practice piece better than other tools but isn’t a complete language education.
Is Speak HIPAA-compliant or appropriate for sensitive content?
Speak is a consumer language-learning product, not a tool designed for sensitive professional contexts. Don’t practice content with private patient or legal information. For most users learning a language for personal or general professional use, the privacy considerations are straightforward.
Is Speak available for my child?
The app’s interface and content assume an older teen or adult user. Kids under ~13 generally do better with Duolingo (free or Super), Little Pim or Muzzy for early childhood, or in-person language tutoring. Older homeschool teens (14-18) studying a Speak-supported language can use Speak productively as part of a broader foreign-language curriculum.
Will Speak replace italki?
No, and shouldn’t try. italki gives you actual human conversation with native speakers, cultural context, real-life accent variation, and the social practice of speaking in front of a person. Speak gives you daily practice volume the human tutor can’t match. Most serious learners use both — Speak daily, italki weekly.
Sources
- Speak — Official site, pricing, languages
- OpenAI Startup Fund blog — “Speak is personalizing language learning with AI”
- TechCrunch — $78M Series C, $1B valuation
- Accel — Connor Zwick interview
- Speak on App Store — iOS app listing
- italki — Human language tutor marketplace
- Pimsleur — Audio-based language learning alternative