Agent builder
Write what the agent should do, pick a voice, upload your policy documents, then call it from the same page with your own microphone. Every setting is versioned, so you can put back yesterday's prompt in one click.
Open the builder →01voice aiworks over · phone · browser · websocket
99ms to first audio · measured19 Indian languages, one voice
Voice agents that answer
in 99 milliseconds
Indic voice AI on your own terms. Nineteen languages in one voice, mid-call language switching, and turn-taking fast enough that callers stop noticing it is a machine. Bring your own phone number.
no card required · measured on our own hardware, under load
latency.cam / turn-timing[synthesised envelope · measured timing]
caller मेरा ऑर्डर कब तक पहुँचेगा?
agent आपका ऑर्डर कल शाम तक पहुँच जाएगा।
agent आपका ऑर्डर कल शाम तक पहुँच जाएगा।
turn gap 99ms·transcribed in 40ms·rtf 0.83
99msto first audiolive call over wss, median of repeated runs
40msto transcribein the live call path, on 8kHz telephony audio
10languages, both waysunderstood and spoken, so a caller can converse. 19 spoken, 11 understood
0.83real-time factorbelow 1.0 is what makes a live call possible
02latency budgetmeasured 2026-09-05
Where the time actually goes
A single headline number is easy to pick favourably. Here is every stage of one turn, in the order the caller experiences it, including the slow one.
The model is the slow part and always will be. The reason a call feels instant is not that 587ms became fast — it is that a cached acknowledgement plays in 1.2ms while the model is still thinking, so the caller never hears the gap.
one turn, end to endmedian
| 0.9ms | number formatting | Indian lakh/crore forms |
| 40ms | speech to text | 8kHz telephony audio |
| 1.2ms | acknowledgement | pre-rendered filler from cache |
| 587ms | the model thinking | median. The one stage that leaves our machines |
| 96ms | to first spoken audio | streaming synthesis, real-time factor 0.83 |
perceived wait 3ms·real turn 724ms
03buildtwo ways in · no-code · websocket
Build it in a form, or in a socket
The same engine either way. The dashboard writes the same configuration the API accepts, so nothing is exclusive to one path.
for the team running the phone line
for the developer wiring it up
One websocket
wss://api.latency.cam/ws/stream # send 20ms frames of 8kHz mulaw configure agent_id, language, voice → transcript.partial # live captions → turn_end # 82ms when certain → audio # 99ms to first frame → interrupted # caller talked over it
04reachdrag the globe
Nineteen languages, one voice
A caller can start in Hindi, ask for Tamil, and it is still the same speaker. Twenty-two languages understood, nineteen spoken, and detection is free because the transcript is written in the language's own script.
HindiBengaliMarathiTeluguTamilGujaratiKannadaMalayalamPunjabiAssameseNepaliSanskritMaithiliBhojpuriMagahiChhattisgarhiBodoDogriIndian English
The points are where Indian-language calls actually happen — the metros, the second cities, and the diaspora markets that ring them. Every one of these is a place a caller may pick up in a language your agent has to follow.
05pillarsfour things that make it sound human
Why it does not sound like a bot
Four behaviours, each one a measurement rather than an adjective.
011.2mscache read
It answers before it has an answer
The instant you stop speaking, a pre-rendered acknowledgement plays from cache in 1.2ms while the real reply is still being written. Measured turn time is 724ms. Perceived wait is 3ms. Silence is what makes an agent feel broken, not latency.
0282msfastest turn end
It knows when you have finished
A silence timer alone either clips people who pause mid-sentence or feels sluggish. A turn-completion model runs 40ms into every pause and ends the turn early when it is certain. On 24% of turns that saves 428ms. When it is unsure, nothing is lost.
03200msto interrupt
You can talk over it
Interruption is detected 200ms into your speech, gated at roughly three words so the "haan", "ji" and "achha" that Indian callers say constantly do not cut the agent off. On a real phone line we also tell the carrier to drop what it has buffered.
040.818voice identity held
One voice, nineteen languages
Switch from Hindi to Tamil mid-sentence and it is still the same speaker. Voices are named per language, so following a caller across languages used to hand them to a different person at every switch. Verified by speaker-similarity scoring, not by ear.
06capacitymeasured, not modelled
0.83 rtf
Synthesis holds a real-time factor of 0.83, so a reply is produced faster than it plays and cannot fall behind the caller however long it speaks. Aggregate throughput measured at 7.59× real time under parallel load. Concurrency is provisioned per account — tell us the number of lines you need.
7.59×aggregate real timeat 16 concurrent streams
0.83rtf, single streambelow 1.0 or the agent falls behind
200msto interruptand not by “haan” or “ji”
0rate limits216 requests, 30-way parallel, zero 429s
what it does
Understands 11 Indian languages
and names which one it heard, with no language code supplied
Converses in 10, speaks 19
preset voices, or clone your own from a recording
Switches language mid-call
follows the caller, or obeys a spoken request
One voice throughout
the agent does not change identity when the language does
Knows when you have finished
prosody, not just a silence timer
Can be interrupted
200ms, and not by "haan" or "ji"
Answers from your documents
an English policy answers a Tamil caller
Brings your own telephony
your carrier, your numbers, your minutes
07known gapswhat does not work yet
The things we have not solved
This is where a page like this usually puts compliance badges. We do not have them yet, and listing what is actually broken is more use to you than a logo.
voice cloning
Tested and rejected. Cross-language similarity scored 0.149 against a 0.213 floor for two unrelated speakers — worse than chance. Holding one stock voice scores 0.818, so that is what ships.
india-hosted calls
Our compute is not in India yet, so audio makes a longer trip than it will once it is. Worth knowing before you integrate, and it is the next thing we are moving.
noisy lines
Speech detection is energy-based. Past about −30 dBFS of background noise it cannot separate speech from the room, and it logs that it cannot rather than failing quietly.
Talk to one in your own voice
Build an agent and call it from your browser in about two minutes. Your phone number, your telephony account, your minutes.