Voice AI · Infrastructure
Bhasha Voice
Voice agents that answer the phone in the languages big AI forgot.
A voice-agent platform for banks, telcos, hospitals and government helplines that handles real calls in Bangla and its regional dialects: understanding, answering, taking actions in back-office systems and handing over to a human when it should.
Launch market: Bangladesh · 5 cited sources · Facts reviewed 14 October 2026
- Cash to MVP (est.)Our estimate of the cash needed to reach a working MVP: one-time setup plus monthly running costs for the build period. It excludes founder salaries.
- $6,370
- MVP timelineMinimum viable product: the smallest version real users will use and pay for, built to learn fast.
- 10 weeks
- Margin after AI (est.)Of every 100 you charge, how much is left after paying for the AI models and infrastructure behind it. Higher means a healthier business.
- ~60%
- Editorial scorePlankito's own judgement: the average of six 1–5 ratings (AI leverage, moat, competition gap, capital efficiency, timing, scalability). It is an opinion to compare ideas, not a forecast.
- 4.3 / 5
At a glance
- Who uses it
- People calling banks, telcos, hospitals and public helplines.
- Who pays
- Enterprises, per handled minute plus a monthly platform fee.
- Who should build it
- Speech and telephony engineers plus an experienced enterprise seller.
- The hardest part
- Dialect accuracy at phone-call speed, and surviving long enterprise sales cycles.
01
Opportunity
Hundreds of millions of people are voice-first. They call rather than type, and they speak languages and dialects that mainstream speech AI handles poorly. Call centres are understaffed, waits are long, and every new service — a bank account, a benefit, a hospital appointment — is gated by a phone queue.
The evidence
95%+
Globalof the world's linguistic diversity is made up of low-resource languages, which mainstream language models underserve.
16% vs 29%
Globalgenerative AI use in the Global South versus the Global North. The gap widened through 2026.
280×
Globalcheaper to run a GPT-3.5-level model than two years earlier — what makes per-minute voice AI economical.
265M+
RegionalBengali speakers worldwide — a single launch language larger than most countries.
Why now
- Real-time speech models have become fast and cheap enough for natural phone conversation.
- Agents can now read and write into back-office systems, so a call can actually resolve something.
- Global voice-AI vendors spread thinly across many languages, leaving depth in any single low-resource language open.
- Banks and telcos face rising call volumes from digital services and need cost relief without losing customers.
02
The AI case
What AI makes possible here, why the big labs won't absorb it, and whether the margins hold.
AI leverage — what is newly possible
- Understands natural, accented, code-switched speech on noisy phone lines in real time.
- Takes actions during the call — check a balance, book an appointment, log a complaint — through system integrations.
- Summarises every call and tags intent, so supervisors see what customers actually ask for.
- Hands over to a human with full context when confidence drops or the caller asks.
Lab-proof moatsA reason competitors, including big AI companies, can't easily copy you: data, distribution, trust or local knowledge.
Dialect speech data
Thousands of hours of consented, labelled call audio across regional dialects. It cannot be scraped from the web.
Back-office integrations
Connectors into core banking, telco billing and hospital systems. Industry analysts note voice agents only work with read/write access to legacy systems.
Regulated-customer trust
Banks and government buy slowly and rarely switch once a vendor passes security review.
Inference economicsThe cost of running AI models each time a customer uses the product. It decides whether an AI business makes money. (estimates)
- Unit
- Minute of handled call
- AI cost per unit
- ~$0.03–0.05
- Price per unit
- ~$0.10–0.12
- Gross margin
- ~60%
How the margin is protected
- →Self-host open-weight speech models fine-tuned on dialect data once volume justifies it.
- →Use a small model for intent routing; call the large model only for complex turns.
- →Price per resolved call, not per minute, once resolution rates are proven.
03
Market & expansion
Launch in Bangladesh. Then expand.
- Start with one vertical where calls are repetitive and costly: bank card services, telco recharge, or hospital appointments.
- Buyers measure average handling time, resolution rate and customer satisfaction. Quote those, not model accuracy.
- Most voice-agent startups sell B2B, and voice becomes a wedge into wider workflow automation.
- Validate: monthly call volume, cost per human-handled call, and the share of calls that are routine.
Editorial scorecardPlankito's own judgement: the average of six 1–5 ratings (AI leverage, moat, competition gap, capital efficiency, timing, scalability). It is an opinion to compare ideas, not a forecast.
- AI leverage
- 5/5
- Moat durability
- 4/5
- Competition gap
- 4/5
- Capital efficiency
- 3/5
- Timing
- 5/5
- Scalability
- 5/5
Where it travels next
India
Dozens of major languages and dialects, with enterprise demand already proven.
Africa
Swahili, Hausa, Yoruba, Amharic and more: huge voice-first populations, few quality speech systems.
Southeast Asia
Regional languages beyond the national ones remain poorly served.
04
Competition & edge
Who already plays here, what they leave open, and what could go wrong.
Global speech & voice AI (e.g. ElevenLabs, Google Cloud Speech)
Multilingual speech recognition and synthesis.
GapBroad but shallow per language; weak on dialects, noisy lines and code-switching.
Indian language AI labs (e.g. Sarvam AI)
Foundation models for Indian languages.
GapFocused on Indian languages and platforms; not building deployed agents for this market's institutions.
Outsourced call centres
Human agents at scale.
GapCostly, hard to scale overnight, inconsistent quality, no instant summaries.
Where the edge comes from
- Depth in one language beats breadth in a hundred for the customers who actually speak it.
- Owning the data pipeline — collection, consent, labelling — compounds with every deployed call.
- Integrations turn a talking demo into a system that resolves problems.
Risks and how to handle them
Global vendors improve fast in Bangla.
Compete on dialect accuracy, integrations and on-premise options for regulated buyers, not raw speech quality alone.
Long enterprise sales cycles.
Start with a paid pilot on one call queue with clear success metrics; expand from proof.
Privacy and recording regulation.
Explicit consent prompts, local data residency and role-based access to recordings from day one.
05
Business model
Usage fees
Per handled minute, moving to per resolved call.
Platform fee
Monthly fee per deployed agent with integrations and analytics.
Speech API
Dialect-tuned speech recognition sold to other developers.
06
Stack & data
Modern, proven tools a small team can run without a DevOps hire.
Telephony
SIP trunking with a local carrier + a real-time media server
Low-latency audio straight from the phone network.
Speech
Streaming speech-to-text and text-to-speech; fine-tuned open-weight models at scale
Latency and dialect accuracy decide whether callers stay on the line.
Agent
Claude API with tool use for reasoning and actions
Reliable multi-turn reasoning and structured tool calls.
Console
Next.js 16 + Tailwind CSS v4
Agent builder, call review and analytics for customers.
Database
PostgreSQL + Drizzle ORM; object storage for audio
Calls, transcripts and audit logs with strict access control.
Evals
Dialect test sets and recorded-call regression suites
Prove accuracy per dialect before every release.
Hosting & ops
Vercel (console) + GPU cloud or local data centre (speech)
Speech workloads need GPUs and, for some buyers, in-country hosting.
Data model6 core tables, for the engineer starting the first migration.
tenants
Customer organisations.
- id
- name
- sector
- data_region
- plan
agents
A configured voice agent.
- id
- tenant_id
- language
- dialects[]
- prompt_version
- tools[]
calls
One phone conversation.
- id
- agent_id
- caller_hash
- started_at
- seconds
- resolved
- handed_off
turns
Utterances within a call.
- id
- call_id
- speaker
- transcript
- audio_offset_ms
- confidence
tool_calls
Actions taken during calls.
- id
- call_id
- tool
- input
- output
- latency_ms
datasets
Consented audio for training.
- id
- dialect
- hours
- consent_basis
- labelled_by
07
Costing
Directional monthly estimates in USD. Your numbers will differ — the structure is what matters.
| Item | MVP / monthWhat the product costs to run each month (hosting, AI models, tools) at MVP stage. | Scale / month |
|---|---|---|
| Telephony & numbers | $80 | $1,500 |
| Speech (STT/TTS) | $150 | $4,000 |
| Model inference | $100 | $2,500 |
| Hosting, DB & storage | $60 | $800 |
| Data collection & labellingConsented dialect audio | $400 | $3,000 |
| Monitoring & analytics | $0 | $150 |
| Total | $790 | $11,950 |
One-time setup
08
MVP & roadmap
Shippable in about 10 weeks, then a path to production.
Build
- One use case (e.g. card blocking or appointment booking) in standard Bangla plus two dialects.
- Inbound phone number, real-time conversation and human handover.
- Two back-office actions through a sandbox integration.
- Call summaries and a review console.
Not yet
- Outbound campaigns
- Self-serve agent builder
- Other languages
- On-premise deployment
Roadmap to production
Validate
Weeks 0–4- Interview 10 call-centre heads
- Record and label 50 hours of consented audio
- Benchmark existing speech models per dialect
MVP
Weeks 5–14- Live pilot on one call queue
- Resolution rate above 60% on routine calls
- Latency under one second per turn
Traction
Months 4–10- Three paying enterprise customers
- Fine-tuned dialect speech model
- Speech API beta
Production
Months 11–20- Second language
- On-premise option for banks
- Series A readiness
09
Start this week
Your first steps. Tick them off as you go.
Metrics that prove it is working
- Share of calls resolved without a human
- Word error rate per dialect
- Average response latency per turn
- Gross margin per handled minute
10
Sources
Every figure on this page links to its source. Facts reviewed 14 October 2026. Costs and market notes without a citation are our estimates.
- 1.Stanford HAI — Mind the (Language) Gap: LLM development in low-resource language contexts (2025)
- 2.Microsoft — The continued state of global AI diffusion in 2026 (2026)
- 3.Stanford HAI — AI Index Report 2025 (2025)
- 4.Dhaka Tribune (citing Ethnologue, 22nd ed.) — Chittagonian, Sylheti ranked among 100 most spoken (2020)
- 5.Andreessen Horowitz — AI Voice Agents: 2025 Update (2025)