User experience
Start, allow the microphone, then speak
The browser may ask for microphone permission on first use. After that, the person speaks naturally and hears short Punjabi responses.
Discussion draft · CAD · taxes excluded
A Punjabi voice service for finding local help.
A voice-first web service that asks for the person’s current location and need, then offers one link from a reviewed list of government and community services. The initial pilot scope is Ontario. Other jurisdictions and Punjabi varieties require their own resource review and evaluation before release.
Current offer
A working private pilot, followed by native-speaker evaluation.
The private pilot is running with voice and text configured for Punjabi, a server-held provider key, a fixed
list of approved links, and password access. In a five-clip test of clean FLEURS Punjabi (India) read speech,
gpt-4o-mini-transcribe had a macro word error rate of 20.5%; one clip was 57.1%. This set is too
small and too clean to estimate real-world reliability. A client pilot therefore starts with native-speaker
testing, safeguarding, accessibility review, and a named person responsible for keeping the service list current.
Proposed pricing
| Phase | Client price | Included work and deliverables |
|---|---|---|
| Closed evidence pilot | CAD $12,000 one-time | Ontario service list, deployed pilot, 150–250 consented utterances from 20–30 client-supplied or separately funded native speakers, model comparison, cost/latency report, no public launch. |
| Production-readiness program | CAD $35,000–$65,000 one-time | Held-out 60-speaker/1,000-utterance evaluation with client-supplied or separately funded participants, including 200 critical cases, WCAG 2.2 AA review, privacy/safeguarding work, and a launch decision. |
| Managed Ontario service | CAD $2,500/month | Available only after launch approval. Weekly crisis-resource review, monthly review of all others, business-hours incident support, reporting, and a provisional allowance of up to 5,000 five-minute-equivalent visits under the current model assumptions. |
| Usage overage | CAD $0.18 per five-minute-equivalent visit | Provisional planning unit including current-model API usage; finalize only after a 100-session cost baseline. Requote if model/provider prices change. |
Pricing phases are cumulative unless a signed statement of work explicitly credits earlier work. The managed fee and overage are provisional until at least 100 representative sessions establish measured cost. One five-minute-equivalent visit assumes two user-speech minutes and 1.5 assistant-speech minutes. No 24/7 SLA is included.
Native-speaker recruitment/compensation, external legal/privacy counsel, formal accessibility certification, insurance, taxes, phone/SMS charges, and new-country resource research are excluded or billed at cost with written approval. Client may supply qualified participants and reviewers.
Link selection and consent
The app only opens links on an approved list
The model requests a category; fixed code supplies the destination. The app checks for a clear yes in a later user turn before opening it. Spoken output still needs hallucination testing.
Recordings and transcripts
The application does not keep recordings or transcript text
The app does not intentionally retain raw audio or transcript text. Production requires a suitable provider agreement and verified retention controls.
Expansion
Adding another province or country
Each expansion needs its own emergency rules, native reviewers, hidden tests, and a named person responsible for keeping the service list current.
Estimated OpenAI API cost
With a 25% planning reserve, the current five-minute estimates are US $0.0389 for the low case, US $0.1020 for the base case, and US $0.3552 for the stress case. At the base figure, 10,000 visits are approximately US $1,020.42. The observed hardware resembles a Vultr plan listed at US $5/month, but the account plan and invoice were not available for confirmation. Treat US $5 as an estimate. Human operations are separate.
Proposed conditions before public use
These draft acceptance criteria must be agreed with the client’s language, safeguarding, privacy, and accessibility leads before testing.
- 100% observed correct crisis routes across at least 200 hidden critical cases.
- Zero observed invented contacts, unsafe advice, wrong-country emergency actions, or unconsented navigation in the held-out evaluation.
- Routine top-1 route accuracy at least 93%; macro-F1 at least 0.92.
- Gurmukhi and Shahmukhi native-speaker sign-off, with limits agreed in advance for script, dialect, age, gender, and noise condition.
- WCAG 2.2 AA, verified resource catalog, privacy review, and accountable human content owner.
Evidence and cost sources
Review the silent quality lab, recorded live proof, and privacy and pilot limits. Model-budget inputs come from the official Realtime 2.1 Mini, GPT-4o Mini Transcribe, and pricing pages; prices can change.
Prepared for discussion by Aadi · aadityalr123@gmail.com · Evidence status dated 2026-08-27.