If this is happening then it must require some amount of CPU and memory use on the device, which even if hidden from the OS can be detected via power use and heat dissipation. It should be possible to indirectly measure a phone's CPU load (including any hidden processors) by filming it in the infrared and/or by attaching it to power and measuring the wattage consumed from the supply.
If this is happening it must also require some network I/O. Something must be phoning home and reporting the data if the data is to be useful.
Note that there is an intrinsic trade-off between the heat dissipation angle and the bandwidth angle. If the device is doing parsing, speech recognition, AI, etc. locally then it must be consuming more energy and dissipating more heat. If the device does these things remotely it must upload more data. Compressed audio is far more bandwidth intensive than digested meta-data.
This seems like something security researchers can and should look into.