Question Data Compression RPCS?

The compression model is only used for integers (where it actually make sense). The WritePackedFloatDelta always send 1 bit if the value equals the baseline, otherwise the full float 32 bits as uint. And because of the floating point format, compressing the 32bit value with the model (that is based on Huffman) will return the same amount (if not more).

If you want to use compression you need to quantise your float first fo integers (up to some precision). But I would say, compressing voice this way is not the best. There is no guarantee of preserving some of the speech characteristic. You need to use a voice compression codec like Speed, Lyra, Opus etc etc that generate very low bandwidth quantised signal and send this as a packet.

Using RPC also is a highly discouraged. They are reliable and that is really not for voice transmission or to be sent at high frequency (every frame in practice). You should use an unreliable channel for that purpose and use the NetworkDriver to send the voice packets.

On top of that you need also to adjust and compensate for latency and packet losses, by caching the received packets and dispatching them at regular intervals (otherwise you will start hearing cracks in the speech)

I would suggest to just use a P2P connection for voice and/or a relay service and use transport directly for the best performance and customisability.