Artoria2e5
commented
1 year ago I suggest you just try it. The earlier versions (non-end-to-end) of LPCNet works using a traditional encoder and quantizer, and the NN is only responsible for decoding: basically reconstruct the audios from a bunch of "features" in a way that's magically better than a conventional decoder. In that case, improvements can be made by changing the decoder without touching the encoder.
The E2E version seems to have NN bells and whistles on the encoding end, so I've got no idea how that works.
When the model changes in LPCNet, this also means the encoded audio bitstreams with prior models become impossible to decode correctly? I have LPCNet working in production for audio files exchange, but I'm unsure if I can upgrade to a newer version without breaking already encoded files.