latency.cam

Indic voice AI for live phone calls. Speech to text, text to speech, spoken language identification, translation and an OpenAI-compatible LLM. One API key, no rate limits.

Loading status…

Languages understood

15 + EN

Speakable live

26+

Detectable

Models in catalog

Alpha, deployed, and benchmarkedFourteen HTTP endpoints and the live-call websocket are serving real inference. Only /v1/clone returns an honest 501. Every latency and accuracy figure in these docs is measured on the live GPU, not estimated — see benchmarks.

Live call latency: 1ms perceived, ~1.35s realA cached filler phrase plays 1ms after the caller stops talking while the real reply is still being produced. Measured per turn: transcription 166ms, LLM 387ms, time-to-first-audio 793ms. TLS is live on api.latency.cam with a real certificate, and the address is an Elastic IP that survives a stop and start, so you can hardcode it.

First request

Check the GPU is up, then send text and get phone-ready audio back.

# 1. is it on?
curl https://latency.cam/api/status

# 2. speak something
curl -X POST https://api.latency.cam/v1/audio/speech \
  -H "Authorization: Bearer $LATENCY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Aapka EMI 15250 rupay ka hai.",
    "language": "hi",
    "format": "mulaw_8000"
  }' --output reply.raw

Full quickstart →

Point your AI assistant at it

The docs are machine-readable. Give your assistant one of these URLs and it can write correct client code, pick the right model for a language, and explain the caveats without being told anything else.

URL	What it is
/llms-full.txt	Everything in one plain-text file. Paste into any chatbot. Start here.
/openapi.json	OpenAPI 3.1. ChatGPT, Claude and Cursor read this natively.
/api/models	Live catalog with licence risk, streaming support and caveats per model.
/llms.txt	Short index for crawlers.

How to use these with a chatbot →

Endpoints

Endpoint	Does	Status
`WEBSOCKET /ws/stream`	Full-duplex live call	alpha
`POST /v1/audio/speech`	Text to speech	alpha
`POST /v1/audio/transcriptions`	Speech to text	alpha
`POST /v1/translate`	Translate text across 23 languages	alpha
`GET /v1/translate/languages`	Which languages translation supports	alpha
`POST /v1/lid`	Which Indian language is this	alpha
`POST /v1/chat/completions`	LLM, OpenAI compatible	alpha
`POST /v1/normalize`	Make numbers speakable	alpha
`GET /v1/fillers`	Pre-rendered phrases, served in about 1ms	alpha
`GET /v1/models`	Everything available, with the caveats	alpha
`GET /health`	Readiness, no key needed	alpha
`POST /v1/clone`	Voice cloning — not implemented	planned

Full endpoint reference →

The three things that matter most

Normalise your numbers

"₹15,250" read as digits is the biggest robotic tell in India. Leave normalize: true on. Bigger impact than any model choice.

Keep replies under 15 words

Long sentences sound robotic in every TTS engine ever built. Put the limit in your system prompt.

Record voices through a phone

A 24 kHz studio recording squeezed into an 8 kHz phone line sounds fake. Record your voice actor on a real call.

Known gaps

We publish what does not work. Read the gaps →