What 9,000 Learners Taught Us About Daily Speaking Practice (Real Data)

On June 9, 2026 we pulled the anonymized usage data on every learner who has ever signed up for TalkToDia — 9,071 people, 7,383 of whom started at least one conversation — and asked one question: what actually separates the people who keep practicing from the people who quit? This post publishes the real numbers, including the ones that embarrass us. No "studies show," no rounding up. Aggregates only, no individual data.

The headline: the first day decides almost everything, voice beats text, and a daily ritual beats both. Here's the evidence.

How many people actually stick with language practice?

Brutally few, and we'd rather print it than pretend. Of everyone who ever started a conversation with our AI tutors, only about 1 in 10 was back the next day, and roughly 1 in 6 did anything in the first week after day zero. The median learner's entire journey is one day long.

Before you conclude that's an indictment of us specifically (some of it surely is): the broader industry doesn't publish these curves, and the few public datapoints on app-based learning suggest single-digit long-term retention is the norm, not the exception. Language learning has a wanting problem — millions start; the habit is the rare part. Which makes the interesting question: what did the minority who stayed do differently on day one?

What predicts who keeps going? (The depth gradient)

The strongest signal in our entire dataset is how deep the first day's conversation went:

Conversation turns on day 0	Came back within a week
1–4 turns	8.1%
5–9 turns	11.8%
10–19 turns	19.4%
20–49 turns	39.9%
50+ turns (only 33 people)	74.2%

A learner who pushed past ~20 exchanges on their first day was five times more likely to return than one who stopped at three. The honest caveat — this is correlation, not proof: motivated people both talk longer and come back, and we can't fully separate the two. But the practical advice falls out either way, because the behavior is the one thing you control: on day one, with any method, don't sample — converse. Push past the polite opening exchanges into an actual conversation. The learners who treated their first session like a real talk, not a demo, are disproportionately the ones still practicing weeks later.

Does voice practice really make a difference?

In our data, yes — the clearest feature-level effect we have. Learners who ever did a voice call returned in week one at 27.2%, versus 17.5% for text-only chatters — about 1.7× — and the gap persists into days 8–30 (12.4% vs 7.1%). Same caveat as above (callers self-select as bolder), but it matches the pedagogy: speaking aloud is the production effect plus real listening practice, and it's the part most learners under-train. One wince-worthy detail from the same pull: the median voice call lasts exactly as long as our free time allowance — people are mid-sentence when the meter runs out. We're fixing that; the data made it unavoidable.

What does a streak actually do?

Learners who used our daily check-in — a one-tap "I showed up today" — returned at 32.9% in week one and 23.6% in weeks 2–4: roughly 2–3× the baseline, the strongest correlation of any behavior we track. Again partly self-selection — but it agrees with the best-replicated finding in memory research: spaced, repeated practice beats massed practice for long-term retention (Cepeda et al. 2008). The ritual matters more than its size. Five minutes that reliably happens beats an hour that happens twice.

Which learners surprised us?

• German- and French-target learners retain best (~28% week-one return); English learners retain worst (~18%) — even though learning English is our biggest segment. Our best guess: many English learners arrive with urgent, externally-imposed goals and price sensitivity, churn fast if the fit isn't immediate; German/French learners skew hobbyist with intrinsic motivation. We don't fully know, and we're saying so.
• Our own AI talks too much. A same-day analysis of recent conversations found 84.5% of our AI's messages contain a question, and the median AI message is five times longer than the median learner reply. That's an interrogation, not a conversation, and it's our fault, not yours — the prompts literally told the model to always keep asking. We're rebalancing it. If an AI partner ever exhausted you, the data says you were right.

What should a learner actually do with these numbers?

1. Make day one deep. 20+ real exchanges, today, not "I'll explore the app." Depth on day one is the strongest flag we can measure for still-practicing-next-month.
2. Speak aloud early. Voice users in our data stick at 1.7× the rate — and the production research says the words you've spoken are the words you keep.
3. Anchor a tiny daily ritual. The 3× streak correlation + the spacing literature point the same direction: schedule beats willpower, and small-daily beats big-rarely.
4. Past day one, fight the plateau with output, not more input — every pattern in our data rhymes with that advice.

Method note: all figures computed June 9, 2026 from production aggregates (signups, conversation logs, call logs, check-ins); no message content was read for this analysis beyond automated counts; cohorts span August 2025–June 2026. Where subgroups are small (the 50+ turn group is 33 people) we've said so. We'll re-run this annually — if the numbers move, you'll see the real ones.