Build a Hinglish Voice IVR for a Clinic: Weekend Appointment Booking

A 4-doctor clinic in Indore was missing roughly 18 calls a day. The single receptionist could not answer the phone, check the doctor's calendar, and handle a walk-in at the same time. Each missed call was often a missed appointment. We built a voice IVR that picks up every call, talks to the patient in Hinglish, finds an open slot, and writes the booking straight into Google Calendar — and we built it over a weekend. This walkthrough gives you the call flow, the Hinglish prompt design, the slot-filling logic, the calendar write-back, and the three gotchas that will trip you up. ## The Answer in 60 Words Use Twilio to answer the call and stream audio, Whisper (or an Indic ASR) to transcribe Hinglish speech, Claude to run the conversation and decide the next question, and a TTS voice to speak back. The bot collects name, doctor, and preferred time through slot-filling, checks Google Calendar for openings, books the slot, and reads back a confirmation. Weekend build, modest per-call cost. ## Why This Matters Now Small Indian clinics live and die on the phone. Patients call to book, reschedule, and ask "kya doctor available hai aaj?" — and they switch between Hindi and English mid-sentence without thinking. A receptionist who steps away for five minutes loses bookings to voicemail nobody checks. By mid-2025, speech-to-text on Indian-accented Hinglish had become good enough, and Claude's conversation handling reliable enough, that a small clinic could run a booking IVR that actually books — not the rigid "press 1 for appointments" trees patients hang up on. This is the same voice-AI stack we use on TalkDrill, our in-house English-speaking app, where Indian-accent speech recognition is the core problem we solve every day. ## What You'll Need ## The Five Pieces of the Stack ## The Call Flow ## The DIY Walkthrough (A Weekend, Step by Step) ## Designing Hinglish Prompts (The Part People Get Wrong) A few rules that worked for us. Keep each bot turn to one question. Confirm numbers and names by reading them back, because that is where ASR errors hide. Offer two concrete slots instead of an open "kab aana chahenge?" — a closed choice is faster to fill and easier to transcribe. And keep a small lexicon of the doctors' names and common medicine terms so the ASR doesn't mangle "Dr Bhatnagar" into "doctor but nagar." We learned most of this the hard way building Indic voice features. Our deep-dive on a TalkDrill Hindi voice bot and a pronunciation lexicon that cut WER 31% covers the lexicon trick in detail, and our latency work on a 740ms voice round-trip on Indian 4G explains why every 200ms of delay makes a phone bot feel broken. ## The 3 Gotchas Latency makes or breaks the feel. A 4-second pause after the patient speaks reads as "the line dropped," and they hang up. Stream audio instead of waiting for the full utterance, use a fast model for the conversation step, and pre-warm your TTS. Aim for under 1.5 seconds from end-of-speech to start-of-reply. Numbers and dates are where ASR fails. "Teen baje" vs "teen tareekh," "10:30" heard as "10:13" — these errors book the wrong slot. Always read the date and time back and require a confirmation before writing anything. We covered the same trap in our Tally CA-helpdesk IVR built in 2 days. Calendar race conditions double-book. Two patients call at the same minute and both get offered 10:30. The fix: do a final freebusy check inside the write transaction, and if the slot just filled, apologize and offer the next one. Don't trust the availability you read 20 seconds ago. ## A Cost Breakdown Per Call At roughly ₹6–9 a completed call, a clinic that recovers even five missed bookings a day pays for the entire system many times over. The exact figure depends on your ASR and TTS vendors and average call length; benchmark with your own traffic. ## Common Mistakes (When Not to Build This) Your call volume is tiny. If the clinic gets 10 calls a day and the receptionist handles them fine, a voice IVR is solving a problem you don't have. Build it when calls are genuinely being missed. You need rich clinical conversation. A booking bot books. If you need symptom triage or insurance pre-auth over voice, that is a much bigger, regulated build — not a weekend project. Keep the first version to scheduling. You skip the human handoff. A bot with no escape hatch frustrates exactly the patients you most want to keep. Wire the handoff in step one, not as an afterthought. ## Real Example: The Indore Clinic The 4-doctor Indore clinic we built this for was losing roughly 18 calls a day to a single overworked receptionist. After the weekend build, the IVR answered every call, booked the straightforward appointments in Hinglish, and forwarded anything complex. The receptionist's phone load dropped enough that she could focus on patients at the desk. The same voice pipeline runs in production on TalkDrill, our in-house English-speaking app, where the hard problem is exactly this: understanding Indian-accented, code-switched speech in real time and replying fast enough that a conversation feels natural. Everything we learned there about ASR lexicons, streaming, and latency budgets went straight into this clinic build. That is the advantage of building voice features for our own product first — we hit the rough edges on our own users before a client ever does. For a text-channel version of the same idea, see our patient booking portal for a 9-clinic diagnostics group that halved no-shows, and for the WhatsApp route, our multilingual Hindi/English/Tamil chatbot. This whole pattern sits inside our AI and automation practice, and the engineer who'd lead your build is Hrishikesh, our CTO. ## Frequently Asked Questions ### Can a voice IVR really understand Hinglish? Yes, with the right stack. Whisper large-v3 and India-tuned ASR engines handle code-switched Hindi-English well in 2026, especially when you add a lexicon of the clinic's doctor names and common terms. Always read numbers and dates back to the caller to catch the errors that remain. ### How long does it take to build a clinic booking IVR? A focused booking IVR — answer, converse in Hinglish, check a calendar, book, confirm — is a weekend build for an experienced team. Adding triage, payments, or EHR integration extends it significantly. Start with scheduling only and expand once it proves itself. ### What does each IVR call cost to run? Roughly ₹6–9 for a three-minute completed booking call, split across Twilio voice minutes, speech-to-text, the Claude conversation step, and text-to-speech. Your exact cost depends on vendors and call length. Benchmark with your own traffic before committing. ### How do you stop the bot from double-booking a slot? Do a final freebusy check inside the write transaction, not just when you first offer the slot. If two callers race for the same 10:30, the second one's booking fails the recheck and the bot offers the next open slot instead. Never trust availability read seconds earlier. ### What happens if the patient has an emergency? The very first step of the flow detects urgency words in Hindi and English and routes the call to a human or reads the emergency number aloud. A booking bot must never attempt to schedule an emergency. The human handoff is non-negotiable. ### Which is better for Hinglish — Whisper or an Indic ASR? Both work; test on your own audio. Whisper handles code-switching well and is easy to start with. India-tuned engines like Sarvam often edge ahead on Indian names and numbers. Run a small benchmark of real clinic phrases and pick on accuracy and latency together.