In RFC2833, the ATA is listening on the local port for DTMF. When it thinks it senses it, it sends an out-of-band message. Wherever the call enters the PSTN, the corresponding DTMF sound is regenerated and inserted. If the DTMF detection in the PAP2T is buggy, it will mistake speech, generate an out-of-band message to the PSTN gateway, and that gateway will regenerate the tones as instructed.

With inband DTMF, the ATA just passes through the audio unchanged, without itself attempting to detect DTMF, so normal conversation is uninterrupted. The problem is that with various compression schemes, while speech is still intelligible, DTMF is sufficiently damaged that an IVR at the other end will have trouble recovering it.