I've been building and testing voice AI agents for a book, and ran into something that surprised me enough to give it its own chapter.

Call a voice agent and press digits on your keypad. In the pipeline I tested, if the agent listens through a speech-to-text model, it hears nothing: across 6 real calls of nothing but touch-tone keypresses, the transcriber produced zero transcript and never even triggered voice-activity detection. Not "got it wrong," got literally nothing.

Then I ran the same tone-decoder (a Goertzel filter over 20ms blocks, the actual DTMF detection algorithm) against a normal spoken sentence instead of tones. It hallucinated 109 phantom keypresses that nobody ever pressed.

The robust fix for our SIP/RTP path was RFC 4733: carry DTMF as named RTP telephone events instead of asking a speech model or an audio Goertzel detector to infer keypad input. Once I wired that up on a real LiveKit + Twilio SIP call, the same kind of calls that produced zero transcript for tones instead surfaced correct keypresses through room.on("sip_dtmf_received"), no STT model involved, 12/12 correct across real calls.

Full writeup with the actual traces is chapter 13/14 of "Building Production Voice AI Agents" (https://www.amazon.com/dp/B0HL53Y5HK), but this repo and its tests stand on their own.