Including removing useless code

Removing one of the 2d conv layers for pitch estimation reduces
complexity without noticeable degradation. FARGAN model has more
adversarial training.
Also, no need for the double precision in the low-pass filter.

Remove old pitch estimator and retrain all models

We're back to processing 10 ms at a time and have no need for 40-ms
"superframes".

Add arch to the LPCNet states, but right now it's always set to zero

also leading underscore in _lpcnet_lpc() wasn't a good idea

That part was not bit-exact

Packet loss concealment based on LPCNet (work in progress)

Jean-Marc Valin authored 3 years ago and Jean-Marc Valin committed 3 years ago

Down to 20 features

Would cause unused pitch embedding vectors to be used

preventing the pitch from going above 255