Speak

Speak started with one hard technical niche: understanding heavily accented speech. Incumbent speech models had about an 80% error rate on learners' accented speech, so Duolingo and Rosetta Stone offered little speaking practice. Connor Zwick and Andrew trained their own model over a weekend on scraped YouTube data and beat state-of-the-art accent detection records.

They saw the wedge as the start of a flywheel: distribution brings data, data and compute bring better models, toward a full end-to-end tutor they expected to be possible in about 10 years. Pronunciation matters most in tonal languages like Mandarin, where the wrong tone changes the meaning.

Sub Club▶ 11:11

Test with users alone; Speak found that learners speak freely without a judge. Early Bay Area testers with weak English behaved very differently when Speak's team sat in the room than when they used the product on their own. Alone, they stopped editing themselves for fear of sounding silly and loved speaking out loud.

The team first had to check whether the model was accurate enough to feel pleasant, and only then tested how the experience felt. "If we were in the room testing with them, it was very different than if we gave them the product to try on their own."

Sub Club▶ 54:33

A phoneme recognizer giving per-sound pronunciation feedback strengthened the whole funnel. Speak's first phone recognizer understood the actual sounds a learner made, so it could coach pronunciation down to the individual phone. On launch, "the entire funnel get substantially stronger."

Connor describes Speak's growth as a series of step functions tied to core model changes, not a smooth exponential curve, and says these moved metrics more than interface tweaks.

Sub Club▶ 58:00

Launch open-ended features only once models can deliver them well. For its first years Speak's language modeling was weak, so lessons stayed rigid and structured. When LLMs improved, Speak launched roleplay, where learners practice any scenario, and it completely changed long-term retention and time spent in the app.

Speak was early to roleplay, built a suite of features around it, and is now best known for it. Better models also let Speak localize quickly, teaching English to speakers of 12+ first languages and adding Spanish, Italian, French, Japanese and Korean.

Sub Club▶ 38:50

Speak picked the hardest, best-served market to prove a 10x product. After market research and travel, Speak chose South Korea, where English spending is cited at 1% of GDP and Gangnam skyscrapers are full of English classrooms. It spent about four years only on Korean speakers learning English, a market neither cofounder had visited before.

Korean learners care only about efficacy, so the crowded market gave the clearest signal on whether the product worked. "The only way we would be able to rise above the noise in a market like that would be if the product was actually truly differentiated."

Sub Club▶ 21:41

Treat entering a country as brand building, not install buying. In Korea Speak combined PR and earned media, owned social, influencers and performance marketing, with organic channels feeding the paid flywheel. It became the number one English education brand there, with 30 to 40% top-of-mind awareness, 70 to 80% aided awareness and about 10% of South Koreans having tried it.

The same integrated playbook was then copied into Japan, Taiwan, the US and Europe. "In Korea we really think about it as a brand building exercise."

Sub Club▶ 25:37

Expand only after fit is solid, then use LLMs to localize fast. Speak opened Japan only once product market fit, pedagogy and platform were stable, which coincided with LLMs making localization and curriculum building far faster than the large upfront investment Korea had needed. Going from one market to two exposed unexpected problems, which Speak automated before moving into many markets within a year.

After Japan it expanded to teaching English to 12+ first languages and added five learnable languages. Japan, Taiwan and the US launched after Korea; Speak later raised a Series C at a $1B valuation.

Sub Club▶ 51:36

Hire a local head of market on launch day, then give autonomy. Speak did not run Japan from San Francisco; it had a head of Japan hired on the day Japan launched and built a local team right away. The central team provides a framework and local teams operate on their own so the brand feels native.

Korea had effectively been an international expansion from the start, so Speak had no naivety left by the time it opened Japan.

Sub Club▶ 20:46

Design for failure cases, since the model works 90 to 99% of the time. Speak has no human in the loop checking the experience, so reliability has to be built into the product itself. That changes how you test, how you plan for failures, and how you separate a visible failure from a hallucination you cannot detect live.

Connor calls it more a product challenge than a technical one, and a defining one since the beginning: telling apart "what's possible versus what seems possible with a demo."

Sub Club▶ 43:03

Over-moderating an AI tutor became a bigger risk than under-moderating. Speak put GPT-4 in production three or four months before ChatGPT launched, testing with 30 power users; one averaged about 8 hours a day for 3 days, using the tutor as a simulated boyfriend. Speak added heavy moderation to avoid a viral inappropriate screenshot, then dialed it back when users were offended at being censored.

The team first suspected broken logs when it saw the 8-hour usage. Users had very specific scenarios they wanted to practice, and censorship got in the way.

Sub Club▶ 49:15

Tell teams not to hold back features over AI costs. Speak assumes model costs will keep falling, so any feature showing ROI should be built without cost-based self-constraint. Its premium subscription leaves margin room, and its token use per learner is lower than that of workflow or coding tools.

Multimodal speech is still expensive, but Speak has far less margin pressure than agent coding companies, which lets it stay on the newest models. Connor says the most cost pressure falls on tools where usage per token is "extremely correlated with value."

Sub Club▶ 1:02:00

Speak bets on relevance and relationship over gamification to motivate learners. Speak believes a highly personalized experience and a sense of relationship with the AI motivate more than streaks and dopamine loops; its aha moment is feeling able to speak out loud. It contrasts this with Duolingo users holding 700-day streaks who say they cannot speak the language.

Its next goal is a tutor that does mastery evaluation, pedagogical planning and dynamic curriculum, inspired by Bloom's 2 sigma effect of one-on-one mastery learning; Connor expects these to be solved in 3 to 5 years, stated as belief. No outcome or fluency data is given to back Speak's efficacy claims.

Sub Club▶ 15:44

Paid acquisition worked almost only on Facebook for Speak's older audience. Speak's Leon Cass reported success nearly only on Facebook, the same pattern as Simply, and tied it to older users needing more time to warm up than app-native millennials.

Subversive

Connected workflow tools help Speak scale while staying lean. Connor aims to keep few nodes and low communication tax as Speak grows, saying AI means teams need not be as big. Connected tools like Notion and Linear make it far easier for new hires to get context than a year earlier.

Sub Club▶ 30:30