# latency.cam

> Indic voice AI API for live phone calls. Speech to text, text to speech in 38
> voices, spoken language identification, translation, and an OpenAI-compatible
> LLM. Built for 8kHz Indian-language telephony.

One API key. No rate limits, no quotas, no per-model gating, no concurrency caps.

## What is live right now

Fourteen HTTP endpoints and one websocket, all serving real inference:
- WEBSOCKET wss://api.latency.cam/ws/stream — full-duplex live call, everything server side
- POST /v1/audio/speech — 38 voices, 19 languages, streams
- POST /v1/audio/transcriptions — 22 languages understood
- POST /v1/translate — 23 languages, IndicTrans2 on CPU
- GET  /v1/translate/languages
- POST /v1/lid — 42 Indian languages identified from audio
- POST /v1/chat/completions — OpenAI compatible, Bedrock backed
- POST /v1/normalize — makes numbers speakable
- GET  /v1/fillers, /v1/fillers/{id}, /v1/fillers/pick/{language}, POST /v1/fillers/generate
- GET  /v1/models — the catalog, with measured error rates
- GET  /health — no key needed

Not built: POST /v1/clone returns 501 with an explanation. Use the 38 preset voices.

## The one thing to understand about latency

A voice bot feels broken because of silence, not because of latency. A cached
filler phrase plays 1ms after the caller stops talking, while the real reply is
still being produced. Measured turn latency is about 1.5 seconds; perceived wait
is about 1 millisecond. The websocket does this automatically.

## Docs

- [Full reference for AI assistants](https://latency.cam/llms-full.txt): everything in one file, paste this into a chat
- [OpenAPI spec](https://latency.cam/openapi.json)
- [Quickstart](https://latency.cam/docs/quickstart)
- [Live calls and the websocket protocol](https://latency.cam/docs/streaming)
- [Models, with measured accuracy](https://latency.cam/docs/models)
- [Languages](https://latency.cam/docs/languages)
- [Measured benchmarks](https://latency.cam/docs/benchmarks)
- [Limits, and what actually constrains you](https://latency.cam/docs/limits)
- [GPU control](https://latency.cam/docs/gpu)