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

Making LPC computation and prediction differentiable

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

Down to 20 features

Adapted from PR: https://github.com/mozilla/LPCNet/pull/134
by zhuxiaoxu <zhuxiaoxu@ainirobot.com>
but had to be reworked due to previous weight quantization changes.

Avoids having to compute a sigmoid

Saves on the MDense/softmax computation since we only need to compute
8 values instead of 256.

Would cause unused pitch embedding vectors to be used

I guess it's a bad idea to forward inputs directly

David Rowe authored 6 years ago and Jean-Marc Valin committed 6 years ago

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

Without the 0.1 bias, the rounding error could cause an offset of -1

Still need to properly dump as sparse.

Completes optimizations from Section 3.6 of the LPCNet paper.