Indie AI voice platform achieves 700ms latency with custom full stack for India

--demigod-- · reddit · 2026-08-21

An AI voice agent platform tailored for the Indian market claims 700ms latency and 2/min costs by running its own models (LLM, TTS, ASR) and owning the carrier layer instead of relaying to OpenAI/ElevenLabs or renting SIP trunks. It uses a 30B parameter model by default for viability but can switch to larger models for complex tasks. The author invites testing on pronunciation, barge-in handling, and naturalness across Hindi, Tamil, Telugu, and Bengali.

Original post →

More from coding & agent

coding & agent channel →