Speech-to-text, text-to-speech and configurable voice agents, built for Indic languages and the English–Indic mixing that real conversations are full of. One API, one dashboard, latency you can measure per turn.
$ curl https://api.voxai.dev/api/v1/stt/transcribe
-H "X-API-Key: $VOX_API_KEY"
-F "file=@call.wav"
-F 'options={"language":"hi-IN","diarize":true}'
{
"text": "Namaste, main apne order ke baare mein…",
"language": "hi-IN",
"speaker_count": 2,
"latency_ms": 312
}11
Indian languages
plus English code-mixing
<800ms
Target turn latency
end of speech to first audio
3
Products, one API
speech, voice, agents
ap-south-1
Data residency
DPDP aligned
Three products
Each product ships with a playground in the dashboard and a documented API. The agent runtime is the three of them wired together with turn detection, tool calling and telephony.
Streaming and batch transcription with speaker diarization, word-level timestamps, automatic punctuation and per-word confidence. Custom vocabulary boosts the product names and account numbers that generic models mangle.
A multi-voice library across languages, genders and speaking styles, with prosody control over pace, pitch and emphasis. Streaming synthesis is optimised for time-to-first-byte, because that is the number a caller actually feels.
Speech, reasoning and synthesis wired into one runtime with barge-in, tool calling and live transfer. Every call is recorded with a per-turn latency breakdown, so a slow conversation is a debuggable one.
Half a support call is Hindi, half is English, and the product names are neither. Models are selected and evaluated on code-mixed audio rather than clean read speech.
The free tier includes an hour of transcription, a hundred thousand characters of speech and a full agent workspace. No card required.