3. Feature-by-Feature Scrutiny (Your Stated Features)
Categories: TS = Table stakes · D = Differentiator · X = Distraction · L = Liability.
3.1 Classification
| # | Your feature | Category | Reasoning |
|---|---|---|---|
| 1 | Start from the learner's current level, not from zero | TS (reading placement) → D (speaking placement) | Duolingo already has an adaptive placement test and "jump here" tests, and Babbel and Busuu have level checks. As a pitch, "we don't start from zero" is weaker than it feels. It becomes a differentiator only if placement is done by speaking: describe a picture, answer 3 questions, read 2 sentences. It should score fluency, vocabulary range and pronunciation, and show the learner why they landed where they did. That's the placement no free app offers. |
| 2 | Curated by goal: official communication vs casual conversation | D (official / work) · X (casual as a paid track) | "Official" has a buyer with a deadline: interviews, customer calls, meetings, hotel front desk, retail floor. "Casual" is a vitamin, and Duolingo gives it away free. Make work the paid core, and keep casual scenes as free warm-ups. Go further than "official": pick situations (interview, customer complaint call, check-in, stand-up), because that's how learners describe their problem. |
| 3 | Several languages: English, Hindi, Bengali | X as three target languages · D as explanation languages | Three target curricula means three level frameworks, three assessment problems and three audiences. There is also no CEFR-style framework for Hindi or Bengali as a second language, and no commercial pronunciation API scores Bengali. Proven payment is in learning English. Use Hindi and Bengali as the languages the app explains in. That's cheap, since the scenes are shared, and it's the proven SpeakX pattern. Learning Hindi or Bengali as a new language can wait for evidence. |
| 4 | Gamified city map with scenes (park, hotel, shop, a stranger asking directions, office) | TS / D (if the places are real work situations) · L (build cost if it becomes a game) | Duolingo Adventures and free Speaking Adventures already put learners in scenes with characters. The map is navigation, not a moat. It earns its keep when (a) the places are the learner's real high-stakes situations, (b) places unlock with level, which makes progress visible, and (c) it's a 2D illustrated map with tap zones, built in days. A game engine, open world or avatars costs months and changes nothing about renewal. |
| 5 | Challenge formats: typing, choosing the right answer, talking | TS | Every app has them. Typing and multiple choice are cheap and good for warm-ups. Talking should be most of the paid time, because speaking is the unmet need. |
| 6 | A voice companion to practise with daily | D (weak alone → strong with a score) · L (persona, data) | AI voice conversation is now in Duolingo, Babbel, Memrise, Speak, Praktika, SpeakX, Stimuler and free assistants. It isn't unique. It becomes a differentiator with: (a) a fixed rubric and a weekly speaking score with before/after recordings, (b) corrections in Hindi or Bengali, (c) scene memory ("last time you rushed the greeting"). Liabilities: a "friend" persona that asks about feelings pulls it into New York and California companion-chatbot laws. Stored voice recordings are biometric-data exposure. Keep it a coach that talks about the scene, and delete audio after scoring. |
| 7 | More human-like voice as models improve | TS (soon) | Everyone gets the same models at the same time. Voice quality will be table stakes, not an edge. It matters for Bengali, where speech recognition and text-to-speech are measurably weaker than Hindi (word error rates ~17–25% vs ~8–15% on public benchmarks). Test Bengali quality before promising it. |
| 8 | Save conversation lessons and reuse them for other users with tweaks | D (for unit economics), if designed as "scripted spine, live edges" | The right instinct, applied to the wrong layer. What can be reused: scene scripts, branches for each level, NPC opening lines, hints, model answers, typing and multiple-choice items, and all NPC audio (generate TTS once, then serve the file). That's most of the content. What can't: the live reply to what this learner just said, and their feedback. Free-form turns differ every time. Duolingo does the same: a fixed scaffold with live turns, and it has pushed the live part below $0.01 a call using open models. So cache the spine, make the live edges short and cheap, and cap live minutes per plan. |
| 9 | "Many gamification ideas" | X (risk) | Streaks, one scene a day and places unlocking by level are enough. Leagues, avatars, coins, shops and cosmetic unlocks are how the build expands to avoid selling. Add a mechanic only when a metric says renewal is the problem and the mechanic addresses it. |
3.2 Scoreboard
| Category | Count | Features |
|---|---|---|
| Table stakes | 3 | Reading-based placement, typing and multiple choice, a better voice over time |
| Differentiator | 5 (conditional) | Speaking placement, work-situation scenes, explanations in Hindi or Bengali, voice coach with a score and memory, reusable scene spine (cost) |
| Distraction | 3 | Three target languages at launch, a paid casual track, open-ended gamification |
| Liability | 3 aspects | A "friend" persona, stored voice recordings, a full game build |
About half the idea can become a reason to choose you, but only after the reframing above. The differentiators come from one place: the app knows how you speak, puts you in your real situations, and shows you getting better. Neither the map nor the voice model is the moat. The speaking data, the work-scene library and the progress record are.
3.3 Feature bloat and building to avoid selling
- Designing the city before anyone pays. The map is the most fun part to build and the least important to sell. A WhatsApp concierge can test the paid core (scene prompt → voice note → feedback) in a week.
- Three languages × two goals × many scenes at launch. Start with 5 work scenes at 3 levels, in English, with Hindi and Bengali explanations.
- A 3D or open-world build. Months of work, and a game engine doesn't change renewal.
- Chasing the most human voice. Use a cheaper, good-enough voice. What learners pay for is feedback accuracy and visible progress.