How it works
The voice AI stack, layer by layer.
A conversational voice bot is six systems in a loop, running under a latency budget of roughly 800 milliseconds. Every vendor in this directory either assembles these for you or hands you some of them. Knowing which is which is most of the buying decision.
The loop runs once per conversational turn — often several times a minute, every one of them inside the same latency budget.
Layer 01
Telephony
Carrier connectivity — a number people can dial, or a trunk you dial out on. Handles ringing, answer detection, DTMF, transfer and hangup, and hands you a live audio stream. In India this layer also carries the regulatory weight: DLT registration, DND scrubbing and consent for outbound.
What you have to build here
- Choose between a CPaaS number and SIP trunking into your existing carrier
- Wire media streaming (WebSocket or SIP media) into your agent runtime
- Handle answering-machine detection — outbound economics collapse without it
- Implement warm transfer to a human, carrying context
- For India: DLT registration, DND scrubbing, and calling-window rules
The mistake people make
Assuming telephony audio is like microphone audio. It is 8 kHz, narrowband and lossy. Every accuracy benchmark you read was almost certainly measured on clean 16 kHz+ audio, and will not survive the trip.
Indicative cost
$0.0120 – $0.1100per minute
India to Philippines, outbound. Rates by country.
Set by the country you are calling and the direction — see the country table. The spread runs better than tenfold, so this is the one layer you cannot estimate globally.
Small per minute, but the one layer with regulatory exposure attached.
Where to read more
Layer 02
Speech to text (STT/ASR)
Streaming transcription of the caller, usually with partial results so the agent can react before the sentence ends. This layer decides more about perceived quality than the language model does: everything downstream reasons over its output, and it cannot recover what the transcript got wrong.
What you have to build here
- Pick a streaming model that handles your languages on telephony-grade audio
- Tune endpointing — how much silence means 'they finished'
- Handle code-switching if your callers mix languages mid-sentence
- Feed a domain vocabulary or hints for names, products and place names
- Measure word error rate on your own recordings, not on the vendor's benchmark
The mistake people make
Benchmarks are quoted on clean read speech. On 8 kHz telephony with background noise and regional accents, real WER is often several times worse. Always test on your own audio.
Indicative cost
$0.0030 – $0.0250per minute
Charged per minute of audio processed. Some providers bill both channels separately.
Runs for the whole call, including silence, unless you gate it.
Layer 03
Language model
Takes the transcript plus your instructions, tools and retrieved knowledge, and produces the next turn. On a voice call it is under a latency budget a chat product never faces — the reply has to start within a few hundred milliseconds or the caller talks over it.
What you have to build here
- Write and version the system prompt — this is the actual product
- Define tools for anything real: lookups, bookings, CRM writes, transfers
- Keep the context small; long prompts cost latency on every single turn
- Stream tokens straight into the speech synthesiser, never wait for a full reply
- Build an evaluation set from real calls before you tune anything
The mistake people make
Choosing a model on reasoning benchmarks. On a call, time-to-first-token matters more than the last few points of quality — a slower, smarter model can produce a worse conversation.
Indicative cost
$0.0020 – $0.0600per minute
Actually billed per token. Converted here to a per-minute figure using typical turn counts — the conversion is where most estimates go wrong.
The layer whose cost you control most directly, by controlling context length.
Where to read more
Layer 04
Text to speech (TTS)
Turns the model's words into audio, ideally streaming the first syllables while the rest is still being generated. Voice quality is what callers actually notice, and time-to-first-audio is a large share of the latency they feel.
What you have to build here
- Pick a voice per language — and listen to it over a phone line, not headphones
- Stream audio out chunk by chunk rather than synthesising whole sentences
- Handle pronunciation of names, amounts, dates and account numbers
- Support barge-in: stop speaking within ~300 ms when the caller interrupts
The mistake people make
Judging a voice on a marketing sample. Studio-quality demos are 24 kHz; your caller hears it downsampled through a phone codec, and some voices fall apart there.
Indicative cost
$0.0040 – $0.0900per minute
Usually billed per character. Only the agent's speech is synthesised — roughly half of talk time.
Often the single largest line item once voice quality matters.
Where to read more
Layer 05
Orchestration and turn-taking
The part nobody budgets for and everybody needs: voice activity detection, endpointing, interruption handling, the state machine, retries, and the real-time transport holding it together. This is where a demo becomes a product.
What you have to build here
- Voice activity detection and endpointing that survive background noise
- Barge-in: cancel in-flight speech and language calls the moment the caller speaks
- A state machine for the call, so it can recover rather than restart
- Backpressure and timeouts on every network hop
- Graceful degradation — what the bot says when a component times out
The mistake people make
Underestimating it. Teams budget for models and discover that turn-taking, interruption and recovery are most of the engineering, and all of the difference between a demo and something you would put in front of a customer.
Indicative cost
$0.0000 – $0.0500per minute
Zero if you build and host it yourself — you pay in engineering instead. Platform vendors fold it into their per-minute rate.
Cheap to buy, expensive to build. This is the real build-versus-buy decision.
Where to read more
Layer 06
Recording, storage and analytics
Call recordings, transcripts, outcome labelling, quality review and the dashboards someone will ask for in week three. Also where consent, retention and redaction obligations land.
What you have to build here
- Store recordings and transcripts with a retention policy you can defend
- Redact personal data — card numbers, identifiers — before anything is stored
- Capture per-call outcomes, or you cannot tell whether any of it worked
- Sample and review calls by hand every week; nothing replaces listening
The mistake people make
Leaving it until later. Without outcome capture from day one, you reach the end of the pilot unable to answer whether it worked, and the pilot gets extended rather than decided.
Indicative cost
$0.0005 – $0.0100per minute
Storage, transfer and any analytics tooling. Small, until retention periods get long.
Rounding error on the bill, load-bearing for the decision.
Speech-to-speech models collapse several of these layers
Some providers now offer a single model that takes audio in and returns audio out, folding recognition, reasoning and synthesis into one call. That removes latency and a lot of orchestration — and removes your ability to see the transcript, swap one layer, or control cost per component.
It is a real trade, not a free upgrade. Ask what happens when you need to log what was said, or change one part of the pipeline.