Voices and models

10 languages, 20 voices, a male and a female in each. Those are the languages an agent can hold a conversation in — understood and speakable. Every voice can be heard before you commit to it.

Hear them first

Voice names tell you nothing about how a voice sounds. "Tamil (Female)" could be anyone. So the agent editor plays them: open an agent, go to the Voice tab, and every voice has a listen button that synthesises a sample through the same engine that will run your calls.

You can also hear your own opening lineType the sentence your agent will actually open with and hear it in each candidate voice. That is a more useful comparison than a stock phrase, because the words you care about are the ones you wrote.

The 19 languages we can speak

Each has a male and a female voice, so 38 in total, and every one of them is reachable on /v1/audio/speech for reading a sentence aloud. Pass the exact voice name, or leave it empty and we pick the female voice for the language.

An agent is offered 10 of these, not 19Speaking a language is only half a conversation. The nine marked below have a voice but no transcription, so an agent set to one of them would talk, the caller would answer, and nothing would come back. They are left out of the agent pickers on purpose. The table underneath this one is the honest version of the count.

Hindi

hi · male and female · agents and speech

Bengali

bn · male and female · agents and speech

Marathi

mr · male and female · agents and speech

Telugu

te · male and female · agents and speech

Tamil

ta · male and female · agents and speech

Gujarati

gu · male and female · agents and speech

Kannada

kn · male and female · agents and speech

Malayalam

ml · male and female · agents and speech

Punjabi

pa · male and female · agents and speech

Assamese

as · male and female · speech only

Nepali

ne · male and female · speech only

Sanskrit

sa · male and female · speech only

Maithili

mai · male and female · speech only

Bhojpuri

bho · male and female · speech only

Magahi

mag · male and female · speech only

Chhattisgarhi

hne · male and female · speech only

Bodo

brx · male and female · speech only

Dogri

doi · male and female · speech only

Indian English

en · male and female · agents and speech

Understanding is wider than speaking

What	How many	Means
Offered to agents	10	Understood and speakable, so a caller can actually converse. Bengali, Gujarati, Hindi, Indian English, Kannada, Malayalam, Marathi, Punjabi, Tamil, Telugu. This is the number to plan against.
Understood	11	Transcribed accurately enough to act on. The ten above plus Odia, which we transcribe but have no voice for, so it is answered in the configured language.
Announcement only	9	Assamese, Bhojpuri, Bodo, Chhattisgarhi, Dogri, Magahi, Maithili, Nepali and Sanskrit have a voice but no transcription. Not offered for agents, because an agent that speaks and cannot listen fails on the caller's first reply. Still available on `/v1/audio/speech` for reading a sentence aloud, where there is nobody to misunderstand.
Detectable	42	Identified from audio alone, without transcribing it.
Translatable	23	Text in and out. English to any of them is direct; Indic to Indic pivots.

Why the gap mattersAn agent can listen in a language it cannot reply in. That is not a bug and it is worth planning for: a caller who speaks one of the three we understand but do not voice will be transcribed correctly and answered in the language you configured. Naming your real languages in allowed_languages is what keeps this predictable.

Choosing a model tier

Three, named for what they are for rather than for whose model is behind them. That is deliberate: the tier you want does not change when we change what implements it, so your configuration does not break when a model is retired or renamed.

Tier	Use it when	Cost
`fast`	The default, and what every latency figure on this site was measured with. Right for scripted support, reminders, confirmations — anything where the agent follows a script rather than deciding something.	Lowest latency
`balanced`	Longer instructions and multi-step reasoning held more reliably. Worth the extra time when the agent has to make a judgement.	Some extra latency
`precise`	Best instruction-following of the three. Needs enabling on your account before it will answer.	Slowest

Set it per agent with llm, or per request on /v1/chat/completions. Measured latency per tier →

Accuracy, measured

What	Result	How it was measured
Language identification on a live call	44 / 44	Real human speech, no language code supplied, at both 16kHz and 8kHz telephony audio. This is why mid-call language switching is on by default.
Answering from your documents	100% in the top 3	A 19-language question set against uploaded documents. Keyword search managed 58% on the same set and 0% on Kannada, Malayalam and Marathi.
Same voice across a language switch	0.818 similarity	A live Telugu → Hindi → Tamil call. Two takes of one real person score 0.747 and two different people score 0.213, so 0.818 is comfortably “the same speaker”.
Ending a turn early when the caller has finished	82ms on 24% of turns	Prosody, not just a silence timer. It can only ever shorten the wait — the timeout still covers every other turn.

Your own voice

The preset voices come from about fifty trained speakers, and that is why they sound mechanical on a phone line — it is the training data, not the engine. So you can supply the speaker instead: upload a recording of a real person and the agent speaks in that voice, in every language we support.

Any file works, including an iPhone voice memo or a video. We take the clearest few seconds, transcribe it ourselves so you type nothing, and let you audition the result through a simulated phone line — the only audition worth trusting, because a voice can sound excellent in a browser and poor on the call.

Voice tab in the agent editor, or POST /v1/voices. The model is version 0.5 and nobody has published a listening test for it, so preview before you rely on it. What we still do not do →