xiph
Quick LPCNet update
Just a few updates around LPCNet. The repo has been transferred from the mozilla organization to xiph. Everything should just keep working (let us know if it doesn't) as usual.
Recently there have been many changes to the code. One set of changes is meant to make the synthesis even more efficient, with an equivalent reduction of around 3.5x in CPU for the same quality. The efficiency improvements are documented in a paper that was recently accepted for ICASSP 2022:
- J.-M. Valin, U. Isik, P. Smaragdis, A. Krishnaswamy, Neural Speech Synthesis on a Shoestring: Improving the Efficiency of LPCNet, Accepted for ICASSP 2022.
A second set of changes makes the LPC part fully-differentiable and should make it possible to train LPCNet end-to-end with any set of input features. The changes (which are optional at training time) are documented in the following pre-print:
- K. Subramani, J.-M. Valin, U. Isik, P. Smaragdis, A. Krishnaswamy, End-to-end LPCNet: A Neural Vocoder With Fully-Differentiable LPC Estimation, arxiv:2106.04129, 2022.
Could you please develop an open-source ASIC (Application Specific Integrated Circuit) just for LPCNet ? Such that the synthesis can run in a cheap IC, which is good for the popularity of LPCNet.
Hey @jmvalin -
Are the changes you're referring to in the start of this thread in the repository now?
Also, is there any formal roadmap for plans you have for this repository?
Thanks!
Yes, all of the changes I mentioned are currently in master. There is no formal roadmap at this point.
Yes, all of the changes I mentioned are currently in master. There is no formal roadmap at this point.
Awesome, thank you for the quick response.
Hi @jmvalin ! Is it correct, that by default any new trained model uses 1)efficiency improvement and 2)differential LPC computation? Also what about models listed there https://media.xiph.org/lpcnet/data/ ? Do they use above improvements if listed after April 2022?
