xiph
post-processing tech on the generated waves
LPCNet performs very well on the features from ground-truth waves. However, when the features are predicted using an end2end model or SPSS model, there are some audible artifacts in the generated waves.
Anyone knows If there are some post-processing tech on the synthesized waves to reduce the artifacts.
Some synthesized samples: e2e_lpcnet_samples_share.zip
For example, from Baidu TTS online demo, it can be found from the spectrogram the audio is enhanced apparently by a post-processing tech to reduce the noise.
@candlewill What about dataset size of "e2e_lpcnet_samples_share.zip". It sounds well. My e2e_demo have some noise, and loss about 3.35 for default parameters.
@candlewill Does LPCNet perform better than Griffin Lim on audio sound quality and audio generation speed ?
@tsungruihon I think the LPCNet (train on your own dataset) is better than Griffin Lim both on quality and speed, especially when some post-processing algorithm could be applied.
@candlewill I think the synthesized sample you shared is very good actually. Probably just a little bit over fitting for LPCNet. By the way, I have the problem of training a stable LPCNet for TTS, I found LPCNet hard to converge, any suggestions ? Thanks
@candlewill have you tried any post processing techs to import sound quality?
