What it costs
Voice AI pricing, with the arithmetic shown.
Nearly every quote you receive covers one layer of a six-layer stack. This works out the rest. Every rate below is editable — put your own quotes in, and watch which layer actually drives your bill.
Where the calls go
Your volume
Rate per minute, by layer (USD)
Entered in USD, displayed in USD. Defaults are indicative bands, not quotes — replace them with real numbers from your shortlist.
Per minute
$0.0448
Per conversation
$0.1344
Per month
$2,688
| Layer | Rate | Billed minutes | Per month | Share |
|---|---|---|---|---|
| Telephony(IN) | $0.0120 | 60,000 | $720 | 27% |
| Speech to text (STT/ASR) | $0.0080 | 60,000 | $480 | 18% |
| Language model | $0.0100 | 60,000 | $600 | 22% |
| Text to speech (TTS) | $0.0150 | 27,000 | $405 | 15% |
| Orchestration and turn-taking | $0.0060 | 60,000 | $360 | 13% |
| Recording, storage and analytics | $0.0020 | 60,000 | $120 | 4% |
| Number rental1 × IN | — | — | $3 | 0% |
Where you call changes the answer. Telephony is priced by destination, and the spread is wider than any other layer — the same minute can differ tenfold between markets. Among the cheapest minutes in the world, and the most regulated. DLT registration, DND scrubbing and calling-window rules apply to outbound, and the compliance work costs more than the minutes do. Numbers require KYC and take days, not minutes, to provision.
Why speech synthesis bills less than the others. You only pay to speak the agent’s side, taken here as 45% of call time. Everything else bills for the whole call, including silence — the assumption most estimates get wrong in the optimistic direction.
The number to ask a vendor for. Total spend divided by conversations handled, for one real month of one real customer. Compare that against your figure above. A gap means one of you is leaving a layer out.
Telephony is priced by country. Nothing else is.
A speech model does not care where the audio came from, so five of the six layers cost roughly the same everywhere. Carriers are the exception. A minute dialled to India costs about $0.0120; the same minute to the Philippines costs about $0.1100 — a 9× difference, on the one line item most estimates treat as a constant.
| Country | Outbound / min | Inbound / min | Number / month | What else it costs you |
|---|---|---|---|---|
| India | $0.0120$0.0060–$0.0250 | $0.0090$0.0040–$0.0200 | $3.00 | Among the cheapest minutes in the world, and the most regulated. DLT registration, DND scrubbing and calling-window rules apply to outbound, and the compliance work costs more than the minutes do. Numbers require KYC and take days, not minutes, to provision. |
| United States | $0.0130$0.0070–$0.0250 | $0.0085$0.0040–$0.0200 | $1.15 | Cheap and easy to provision, but carrier trust is the real cost: unregistered traffic gets labelled 'Spam Likely' and answer rates collapse. Budget for STIR/SHAKEN attestation and branded-caller registration, plus TCPA consent obligations on outbound. |
| Canada | $0.0140$0.0080–$0.0280 | $0.0090$0.0050–$0.0200 | $1.15 | Prices track the US closely. CRTC rules on unsolicited telecommunications are stricter than the US equivalent, and the national DNCL is a hard requirement on outbound. |
| United Kingdom | $0.0300$0.0080–$0.0700 | $0.0100$0.0040–$0.0250 | $1.00 | The landline/mobile split is unusually wide here — mobile termination can be four times a landline minute, so your blended rate depends entirely on your list. Ofcom rules on silent and abandoned calls bite hard on outbound dialling. |
| Germany | $0.0450$0.0080–$0.1100 | $0.0120$0.0050–$0.0300 | $1.50 | Mobile minutes are expensive and landline minutes are not, so the blend matters more than the headline. Numbers require a local address, and GDPR makes recording and retention a design decision rather than a default. |
| United Arab Emirates | $0.0900$0.0500–$0.1800 | $0.0400$0.0200–$0.0900 | $25.00 | A closed telecom market: two licensed operators, VoIP termination restricted, and numbers usually obtained through a local partner rather than a global CPaaS. Assume weeks of lead time and per-minute rates several times the Indian equivalent. |
| Saudi Arabia | $0.0800$0.0400–$0.1600 | $0.0350$0.0150–$0.0800 | $20.00 | Similar shape to the UAE — regulated termination, local partner, long provisioning. CITC registration is required for commercial calling, and unregistered outbound is blocked rather than merely labelled. |
| Singapore | $0.0220$0.0100–$0.0500 | $0.0120$0.0060–$0.0300 | $5.00 | Straightforward to buy and reliable to run. The DNC registry under the PDPA applies to marketing calls, and the penalties are real enough to be a design constraint. |
| Australia | $0.0400$0.0120–$0.0900 | $0.0120$0.0050–$0.0300 | $2.00 | Wide landline/mobile spread, and the Do Not Call Register plus calling-hour restrictions apply to outbound marketing. Numbers require an Australian presence with most carriers. |
| Philippines | $0.1100$0.0600–$0.2000 | $0.0300$0.0120–$0.0700 | $10.00 | One of the most expensive mobile terminations in Asia — often the layer that decides whether outbound is viable at all. Model this before modelling anything else; it can exceed every other stack layer combined. |
| Indonesia | $0.0700$0.0300–$0.1500 | $0.0250$0.0100–$0.0600 | $12.00 | Mobile termination is costly and quality varies sharply by carrier route. Cheap routes exist and tend to be the ones with the audio problems that ruin transcription accuracy. |
| Brazil | $0.0500$0.0200–$0.1200 | $0.0200$0.0080–$0.0500 | $4.00 | Mobile is several times landline, and Anatel's rules on telemarketing — including the 0303 prefix requirement for marketing calls — change what a compliant outbound campaign looks like. |
| Mexico | $0.0350$0.0150–$0.0800 | $0.0150$0.0060–$0.0400 | $3.00 | Reasonable rates by regional standards. REPEP, the national exclusion register, applies to advertising calls, and mobile numbers carry a different rate to landlines. |
| South Africa | $0.0550$0.0250–$0.1200 | $0.0200$0.0080–$0.0500 | $3.00 | Mobile termination dominates the bill, and POPIA gives consumers direct-marketing rights that shape consent capture on outbound. |
| Japan | $0.0750$0.0300–$0.1600 | $0.0200$0.0080–$0.0500 | $8.00 | Expensive mobile minutes and strict number-allocation rules that usually require a local entity. Provisioning is slow enough to belong on the project plan, not the checklist. |
Indicative bands for CPaaS-style pricing, leaning towards mobile termination because voice bots overwhelmingly call mobiles. Volume commitment and bringing your own carrier over SIP move these more than anything in the table.
The five pricing shapes, and what to watch
Per minute
Works when — Usage tracks value, and short calls are cheap.
Watch for — Silence bills the same as speech. Ask whether the meter starts at dial or at answer, and whether it rounds up per call.
Per call
Works when — Simple to forecast, and long calls do not punish you.
Watch for — Check what counts as a call. A voicemail, a two-second hangup and a twelve-minute conversation are often billed identically.
Per qualified lead
Works when — Their incentive matches your outcome, which is rare and valuable.
Watch for — Everything rests on the definition of qualified, and who adjudicates a disagreement. Get it in writing before signing.
Platform fee plus usage
Works when — Usually the honest shape for enterprise deployments.
Watch for — The fee is the part left off the per-minute quote. Always ask for the fee and the rate together.
Per seat
Works when — Familiar, and predictable if headcount is stable.
Watch for — Poor fit for automation — the whole point is handling more conversations without more seats.
The best case, and what it takes
Cheapest per conversation
Short, scripted, single-language calls on a self-serve platform with a small prompt and a fast open model. Little orchestration, no custom voice, minimal retrieval.
Every layer near the bottom of its band
Typical production
Multi-turn, multilingual, retrieval-backed, with recording and human handoff. A branded voice and real compliance handling.
Mid-band across the stack, plus a platform fee
Most expensive
Long calls, premium cloned voices, large context on every turn, a frontier model, dedicated deployment and long retention.
Top of band on synthesis and model, plus infrastructure
Four levers that actually move the bill
Shorten the context
The prompt is re-sent every turn. Trimming it cuts both cost and latency, and latency is what callers feel.
Cut dead air
Every layer except synthesis bills for the whole call. Faster endpointing shortens calls without changing the conversation.
Detect voicemail properly
On outbound, a machine mistaken for a person burns an entire call's cost for no outcome.
Match the voice to the job
A premium cloned voice can cost several times a standard one. On a status-update call, nobody notices.
These defaults are bands, not quotes
Provider pricing changes constantly and varies by volume, region and commitment. The numbers above are indicative ranges to give the arithmetic a starting shape — they are not anybody’s price list, and they will drift.
Use this to understand the structure of the bill, then replace every rate with a written quote. The one figure worth demanding: total spend divided by conversations handled, for one real month of one real customer.
Questions people ask about voice AI pricing
- How much does a voice AI bot cost per minute?
- There is no single figure, because a voice bot is six billable layers and most quotes price one of them. At indicative mid-band rates, calling India, the stack above lands near $0.0448 per minute all-in — telephony, speech recognition, the language model, speech synthesis, orchestration and storage combined, plus any platform fee on top. Change the country, the call length or any single rate and that number moves substantially, which is the point of doing the arithmetic rather than accepting a headline.
- Why is telephony priced differently in each country?
- Telephony is the only layer geography prices. A speech model costs the same whatever country the audio came from; a carrier does not. Outbound to India runs about $0.0120 a minute against roughly $0.1100 to the Philippines — about 9 times the price on the one line item most estimates treat as a constant. Mobile termination, volume commitment and whether you bring your own carrier over SIP move it further.
- What does a per-minute quote from a vendor usually leave out?
- Most often the platform fee, and frequently the difference between billed minutes and talk time. Every layer except speech synthesis bills for the whole call including silence, while synthesis only bills the agent's own speech — roughly half of talk time. Number rental, recording storage and retention are separate again, and are fixed costs, so they hurt most at low volume.
- Which layer of the voice AI stack costs the most?
- It depends on where you are calling. In cheap-termination markets speech synthesis or the language model is usually the largest line once voice quality matters. In expensive-termination markets telephony alone can exceed every other layer combined, which is why it is worth modelling before anything else.
- What single number should I ask a vendor for?
- Total spend divided by conversations handled, for one real month of one real customer. It is the only figure that cannot be assembled selectively. Compare it against your own estimate: a gap means one of you is leaving a layer out.