Sound quality from MEMS microphone on ESP32 via ESP-NOW - is 12bit enough?

I am thinking of making a wireless microphone based on ESP-NOW. The microphone is to be used to record the sound (speech) of a distant person.
Transmitter - MEMS microphone with digital transmission or something based on a regular microphone and MAX9814. The latter will allow the use of popular tie microphones. Here, however, is the first question, will the 12-bit A/D converter in the ESP process the sound in good quality? However, does high quality require 16bit? (I'm talking about speech). I think the ESP will collect 100 16bit samples (200 bytes and send them). Such a packet is short about 3ms when sampling 32k. Here is another question, is 32k enough? A short packet should not introduce visible latency. The receiver will collect the data and send it to the DAC (8 bit in ESP32 is definitely not enough for audio). In your opinion, does it make sense to send (ESP-NOW technology) audio in this way? Is it suitable for real-time operation?

I think calibration is not crucial here - it's not a meter. More relevant is how much 12bit will be sufficient for good quality speech transfer. I say "speech" because for music it's certainly not enough.
My second concern - how exactly does the ESP-NOW communication work? The description of the ESP-NOW function mentions acknowledgement which reminds me of TCP rather than UDP. So if a dispatch command goes, will it be immediately executed? If I set up a 3ms buffer then I can't afford to have ESP negotiate something or have any other delays. It simply has to work as a send over the 'radio' at 433MHz

khoam wrote:

The transmission is UDP, unacknowledged (it's basically WiFi over MAC, no IP) - either MAC unicasts or multicasts are sent. Of course, the devices require 'pairing'. Sending of packets is done immediately.

I thought so too, but the manual says "The sending callback function that can be set to inform the application layer of
transmission success or failure." How is this supposed to be implemented? To do this requires an acknowledgement that is typical of TCP and absent in UDP (I'm leaving out UDP's own communication solutions implemented at a higher layer). Unfortunately, the only thing that can be said about the ESPRESSIF instructions is that they are (good and all). The ESP8266 Technical Reference is 117 pages long. For example, the description of the ATTINA 2313 is 274 pages, and these are, after all, completely different classes of processors.

I understand that, in your opinion, the realisation of a smooth transfer at 512kb/s (32ksample/s) with no delay is as possible?