Voice Agents Handbook
Building production voice AI with LiveKit
by Mahimai Raja J
Voice Agents Handbook
Building production voice AI with LiveKit
by Mahimai Raja J
Nine chapters. Three surfaces: phone, browser, mobile. The patterns from voice agents shipped for plumbers, lawyers, and immigration consultants. The shortest path from “add voice to our product” on a Monday to a working agent answering calls by the weekend.
Drag the book to rotateTouch the book to rotate it
Out now. $9.99 USD on Kindle, paperback available.
The brief “add voice to our product” lands on a backend developer’s desk every week now. What follows is usually three weekends of trial and error: which STT, which LLM, which TTS, how to put it on a phone line, why the agent confirms data the caller never confirmed, why the call drops after thirty seconds, why the build cost six times what the demo predicted. The handbook collapses those weekends into one read.
You start at the LiveKit quickstart and end with a voice agent answering calls for your business: a phone number, a web button, a mobile feature, and the deployment shape that does not crash under load. Nine chapters. Three surfaces.
Every code block runs. Every claim has a number behind it.
You’re a backend developer. Someone on your team (maybe you) decided the product needs voice. You’re comfortable with async Python. You’ve shipped LLM-based features before. You’re trying to figure out what voice actually means in code, infrastructure, and provider choices.
If that’s you, this book skips your first three weekends.
If you’ve never opened a terminal, this isn’t the book for you, and I’ll tell you that up front.
This isn’t a guide to training your own STT or TTS models. It isn’t a deep dive on WebRTC internals. It isn’t about chat-style LLM products with optional voice. It isn’t a comparison of every framework: it picks LiveKit and goes deep.
If you want any of those, you want a different book.
IntroductionFree sample
PART I: FOUNDATIONS
PART II: BUILDING REAL AGENTS
PART III: PRODUCTION
Conclusion
APPENDICES
| Stage | Budget |
|---|---|
| Network, user to server | ~50 ms |
| VAD and turn detection | ~100 ms |
| STT finalization | ~150 ms |
| LLM time to first token | ~300 ms |
| TTS time to first byte | ~150 ms |
| Network, server to user | ~50 ms |
| Total | ~800 ms |
Each term is a real, measurable budget. A 600 ms end-to-end feels instant; 1500 ms feels broken. Every 100 ms you shave from the LLM line is directly audible.
The agent that mishears "drain cleaning" as "drainage exorcism" is funny. The agent that won't stop talking when a frustrated caller is trying to be heard is the one that costs you the customer.
Markdown formatting that helps in chat becomes literal "asterisk asterisk" noise in voice.
Opus is the reviewer; Haiku is the receptionist.
Get your copy
Out now. $9.99 on Kindle.
Pick your local Amazon store.
Prefer print? Buy the paperback on Amazon USA →
Free PDF in your inbox.