The constellation just shows that you have a stable signal being decoded, not WHAT the data stream contains.  In an ideal world you'd just see four tight clusters of data in the four corners of a square regardless of whether you are decoding a control channel or voice.  In practice you may see some slight disturbances and spread of the pattern into an X shape.  You may also notice momentary rotation of the pattern due to phase shifts, but these will (usually) go away as the gardner_costas loop adjusts itself.

There are potentially numerous reasons for not hearing voice, but incorrect command line parameters are the most common.  Make sure you have "-V", "-2" and "-U" specified.  When a tgid is tuned, you should see an indication at the bottom of the terminal screen showing the frequency and optional (phase 2 only) slot number as well as the tgid and tag (if available).  If you're not seeing that, it's not attempting to decode audio.

if you are running my version of the code (github.com/boatbod/op25) you have the option to suppress encrypted audio using the "--nocrypt" command line parameter.  In this instance you should see indication on the terminal screen (bottom right) when encrypted audio is being suppressed.  On Max's version of the code, encrypted audio will be played and just sounds like garbage noise.