facebookresearch / facebookresearch/FlowDec
Can postfilter be optimized with the backbone (DAC) in E2E way?
Open
- Dominant language
- Python
- Stars
- 215
- Forks
- 21
- PR merge metrics
- No merged PRs in 30d
Description
I'm not sure training with simple reconstruction loss already makes the codebook utilization and codec encoder parameters optimal, and the postfilter works more like a patch to improve the voice quality.
Have you tried to optimized the whole network in E2E way to see if there's further gain?
Contributor guide
Assessment
This issue has not been assessed yet.