Unspoken is an NVDA add-on originally developed by Bryan Smart and making use of Camlorn_audio. I have since taken over development and switched it to use Libaudioverse. The goal of Unspoken is to replace spoken information with sounds in a similar manner to Objsounds but with 3d audio rendering.
That is, sounds can be moved vertically as well as horizontally, the panning sounds more natural, and the stereo field is widened.
This add-on works with speaker systems, but the 3d effects require headphones.
If a sound is available, it replaces an equivalent spoken message. The sound will be played at a corresponding point in 3d space on a virtual screen in front of you.
Say all and the Internet do not work as well as they could. In the case of say all, you get no sound. On the Internet, reading by paragraph also does not play sounds. Unspoken provides an options dialog, so you can reconfigure it to speak role names if you want. This cannot be fixed without rewriting NVDA's speech components. Conversation about this rewrite is in progress; here is the NVDA ticket.
There may currently be a bug where Unspoken stops playing sounds until NVDA is restarted. I have not seen it myself.