They are two pivots, the first one requests the second one (and vice-versa) . And also there is a collect DTMF with record_call callflow action . Then, the second pivot plays the previously recorded audio file using the URL with another children pivot module that requests the first one again.
The ecallmgr system_config document has the use_shout = true . On the first pivot the playback terminator is the # sign (on collect DTMF).