AI Voice Receptionist (Hindi)

A real-time voice agent that answers a clinic's phone as its named virtual receptionist and takes a booking end to end — greeting, intent, name capture, appointment slot — entirely in Hindi. Built for the businesses that lose the most to a missed call: clinics, gyms and dealerships whose callers will phone rather than fill in a form.

For a dental clinic, an unanswered phone is a patient who simply calls the next result on Google — and staffing a line through evenings and weekends is not economic at that size. The harder constraint is language: the callers who most need a phone option are the ones least likely to fill in an English web form, so an English-only agent solves nothing. It also has to feel fast; more than about a second of dead air and the caller assumes the line dropped.

We built a voice agent that answers as the clinic's named virtual receptionist, holds the entire conversation in Hindi, and walks the caller through booking a checkup — greeting, intent, name capture, appointment slot. Speech recognition, the language model and speech synthesis are streamed into each other rather than run in sequence, so the agent begins replying while the caller is still finishing their sentence. The demo below is a full unedited call.

Key decisions — and why

• Hindi first, not Hindi as a translation layer

  The whole point of a phone line is reaching people who will not use a form. Running the conversation natively in Hindi — rather than translating an English script — is what makes it usable for the callers who actually need it.

• Streamed pipeline, not sequential

  Transcribe, then think, then speak costs three to four seconds end to end and callers hang up. Streaming each stage into the next keeps the reply near a second, which is the difference between a conversation and a voicemail system.

• A visible way to end the call

  An AI you cannot get out of is worse than no AI. An explicit End the Call control — and a handoff on low confidence — keeps the caller in charge.