Skip to content
Idea Atlas
Sign in

BhashaCity

As described:StopNarrowed:Continue

No as described. Go as work English for Hindi/Bengali-medium learners.

language learning · ² At a blended ~$2.50 a month per payer: ₹199/month and ₹1,199/year plans, plus some ₹399 Pro, at ₹88 = $1. Institute seat licences (from month 9) reduce the number of individual payers needed; see plans/11. · last reviewed (next review by 28 Dec 2026) · 60 cited sources

Save
Share

Start this week

  1. Tell me your hours per week, which of Hindi or Bengali you can create content in, and whether you know a spoken-English trainer (or a BPO voice-and-accent trainer) who would advise.
  2. Write 5 work scenes (interview intro, customer complaint call, hotel check-in, shop-floor query, team stand-up) at 3 levels each, with the trainer.
  3. Run a voice-stack test: Sarvam STT + a small LLM + cached TTS, against Gemini Flash Live, on 20 real learner recordings. Measure accuracy on Indian English and cost per minute.
  4. Post the first 10 "say this, not this" videos, and take the first paid WhatsApp concierge sign-ups.

Checkpoints and kill gates

Decide in advance what counts as on track. Missing a checkpoint means changing the buyer or the offer, not adding features.

  1. Day 30

    ~15 hrs/week

    50 paying ₹149–199 for the 30-day WhatsApp concierge

    Full-time

    50 paying ₹149–199 for the 30-day WhatsApp concierge

    If you miss it

    Change the entry scene once (interview-first or customer-call-first). If it misses twice, stop.

  2. Day 90

    ~15 hrs/week

    150 payers. ≥50% month-2 renewal. Median speaking score up ≥10% in 30 days. Voice and model cost <20% of revenue.

    Full-time

    150 payers. ≥50% month-2 renewal. Median speaking score up ≥10% in 30 days. Voice and model cost <20% of revenue.

    If you miss it

    Cut the product down to the single most-used scene track.

  3. Month 6

    ~15 hrs/week

    $1k MRR. ≥25% of new users from shares or referrals.

    Full-time

    $1k MRR. ≥25% of new users from shares or referrals.

    If you miss it

    Pivot A: Bengali-medium only (less competition). Pivot B: institutes first (B2B seats, with B2C as the funnel).

  4. Month 12

    ~15 hrs/week

    $2.5k MRR, with ≥5 institutes paying. Payer churn ≤12%/mo.

    Full-time

    $2.5k MRR, with ≥5 institutes paying. Payer churn ≤12%/mo.

    If you miss it

    Stop adding features. Fix churn and the share loop. Holding a profitable $2–3k is a valid outcome.

  5. Month 18

    ~15 hrs/week

    $5k MRR

    Full-time

    $5k MRR

    If you miss it

    Put everything into the best channel (content or institutes). Decide whether to test paid acquisition.

  6. Month 24

    ~15 hrs/week

    $8k MRR

    Full-time

    $8k MRR

    If you miss it

    Above $8k with CAC payback under 3 months: consider buying growth for the $10k–20k levels. At $2–8k, run it profitably. Below $1k, stop or pivot.

Do

  1. 1.Make work English the product. The paid places on the map are the interview room, the customer call, the hotel desk, the shop floor and the team meeting. Casual scenes (park, shop, directions) stay as free warm-ups.

    Target: ≥70% of voice minutes are in work scenes

    Why: Job seekers have a deadline and a reason to pay. SpeakX's paying base is mostly Hindi-belt job seekers. "Casual" is a vitamin.

  2. 2.Place by speaking, not by reading. Five minutes: describe a picture, answer 3 questions, read 2 sentences aloud. Place the learner on the map at that level, and show them why.

    Target: ≥60% of signups finish it. ≥70% say the level "feels right".

    Why: Duolingo already places by reading. The unmet need is "I understand English but freeze when I speak".

  3. 3.Build a "scripted spine, live edges" scene engine. Pre-generate each scene's branches, hints and NPC audio once, per level. Call the live model only for off-script replies and feedback.

    Target: Voice and model cost ≤20% of revenue. ≤₹1 per voice minute.

    Why: This is where your caching idea works. Free-form voice replies can't be reused, but most of the scene can.

  4. 4.Explain in the learner's language. The partner speaks English in the scene, and corrections and tips come in Hindi or Bengali (text and optional voice).

    Target: ≥40% of learners turn L1 explanations on

    Why: This is what SpeakX proved in Hindi. Bengali-medium learners are served less (SpeakX has only announced Bengali).

  5. 5.Show progress as a score and a recording. A weekly speaking report covers fluency, filler words and pronunciation (Azure for en-IN), plus your day-1 vs today clip.

    Target: ≥50% of payers open the weekly report

    Why: It's the reason to renew. It's also the most shareable asset.

  6. 6.Keep it to one scene a day. A 5–8 minute daily scene with a streak, and one new place unlocked per level.

    Target: Payers practise ≥4 days a week

    Why: The habit, not the map, is what retains. Short sessions also cap voice cost.

  7. 7.Price for India, honestly. Free: placement, warm-up scenes, 10 voice minutes a week. Plus ₹199/mo or ₹1,199/yr (all work scenes, ~120 voice minutes a month, weekly report). Pro ₹399/mo (mock interviews, 300 minutes). UPI Autopay with pre-debit notice, and one-tap cancellation.

    Target: Refunds and chargebacks <3%. ≥35% of payers on annual by month 12.

    Why: ₹199–299 is the proven band (Stimuler, SpeakX). RBI and CCPA rules require clean mandates.

  8. 8.Launch on web (PWA) + WhatsApp. Add an Android app only once month-2 renewal is proven.

    Target: Store submission after the day-90 checkpoint

    Why: Web checkout avoids the Play fee (15%). A PWA handles the microphone on Android Chrome.

  9. 9.Sell seat licences to institutes from month 9: skilling centres, hotel-management and retail-training institutes, and BPO recruiters, with a trainer dashboard.

    Target: 5 institutes, 500+ seats by month 12

    Why: It raises value per learner without ads. Communication is a hiring filter (Versant is used by large BPOs).

  10. 10.Make outputs shareable. A "my speaking score" card, a before/after clip, and a challenge ("can you handle this customer call?").

    Target: ≥25% of new users from shares or referrals by month 6

    Why: Seekho and Kuku buy growth with ad spend close to their revenue. Shares are the channel you can own.

  11. 11.Build the safety and privacy rails in v1: 18+ only, AI disclosure, no raw audio kept after scoring (default), no voiceprints, and a coach persona rather than a "friend".

    Target: Zero stored voice files beyond 24 hours unless opted in

    Why: DPDP (under-18 parental consent), voice-biometric suits in the US, and companion-chatbot laws all key on these points.

  12. 12.Meter voice use. Fair-use caps per plan, one routing rule (small model first), and cached TTS for scripted lines.

    Target: Average cost per payer ≤₹40/month

    Why: A single heavy learner can use ₹300+ of voice on a ₹199 plan.

Avoid

  1. 1.Teaching three target languages (English, Hindi and Bengali) at launch

    Why: Three curricula, three assessment problems and three audiences. English is where proven payment is. Hindi and Bengali are the explanation languages, not the target languages, until demand says otherwise.

  2. 2.Building the city as a real game (3D, open world, avatars, a game engine)

    Why: It's months of work, and a map is UI, not a moat. Duolingo ships its scene mechanic free. Scene-based language games are small indie businesses (the best recent one grossed ~$350k). A 2D illustrated map with tap zones is enough.

  3. 3.Pitching "we don't start from zero"

    Why: Duolingo, Babbel and others already offer placement tests. The real gap is placement by speaking and content above A2. Say that instead.

  4. 4.A "casual conversation" track as a paid product

    Why: It's a vitamin against a free Duolingo. Keep casual scenes as free warm-ups.

  5. 5.Trying to reuse whole conversations between users

    Why: Every learner reply differs, so live turns can't be replayed. Reuse the scene script, the branches, the hints and the audio, and keep the live part small and cheap.

  6. 6.Premium voice models by default (e.g. full gpt-realtime at ~$0.20–0.45/min)

    Why: At ₹199 a month, voice cost would exceed revenue for daily users. Use the cheaper Indic-capable stacks (Sarvam, Gemini Flash Live, cascades) at ~$0.01–0.05/min.

  7. 7.Users under 18, or school-age learners

    Why: DPDP needs verifiable parental consent for under-18s, and COPPA and California SB 243 add duties for minors. Kids' spoken English (PlanetSpark) is a separate, sales-heavy business.

  8. 8.Storing learner voice recordings, or building voiceprints

    Why: BIPA voiceprint suits (Siri, Teams, Otter, Fireflies), Washington MHMDA and Texas CUBI. Score, then delete.

  9. 9.A companion persona that asks about feelings or builds a "relationship"

    Why: That is what pulls a practice bot into the New York and California companion-chatbot laws. Stay a coach that talks about the scene.

  10. 10.Weekly plans, auto-converting trials, or hard-to-cancel mandates

    Why: India's CCPA names "subscription trap". RBI requires pre-debit notice and easy cancellation. NPCI now has a central portal to cancel mandates.

  11. 11.Promising outcomes ("get a job", "fluent in 30 days", "guaranteed")

    Why: ASCI and consumer-protection exposure, and refund requests when the outcome doesn't happen.

  12. 12.Paid ads before month 6

    Why: The category's leaders spend ~₹1 on ads for each ₹1 of revenue. Until renewal is proven, ads only buy churn.

  13. 13.A native iOS app early

    Why: Low iOS share in the target segment, Apple's in-app purchase rules, and review time. Android (or the PWA) is where these learners are.

  14. 14.Adding game features when a checkpoint fails

    Why: Change the audience (job seekers vs employed staff, Hindi-medium vs Bengali-medium) or the entry scene. Don't add badges, avatars or leagues.