HAESUNG JEON
HAESUNG JEON
@kuaashish Any updates?
@JiaYK could you success to train multi speaker model?
It's sad to hear failure. Maybe small(relatively) batch_bins affect training. I had better result with bigger batch_bins in my experiment(500000 vs 2000000) on finetuning setting. I'm gonna train on multi...
@yurtmete I've not started training. since I'm busy recently.. I'm gonna try KAIST AUDIOBOOK DATASET which has 11 speakers(6 male, 5 female maybe?). It's korean dataset
@wanchichen Thanks! I've read perform_kmeans.sh. if storage_save_mode is set, only features for training kmeans are dumped into storage. and these features are used for kmeans training. and actual kmeans train...
@wanchichen i was able to train k-means using MiniBatchKMeans with >50GB features, on 32GB ram machine using numpy mmap. (it took long time)
이해가 안되는 부분이 있어서 질문 남깁니다. Forward Probability와 Backward Probability를 정의할 때 alpha_t(s), beta_t(s) 이런식으로 정의하는데, s가 l에 대해서 쓰여져 있습니다. 수식5와 9에서 l의 밑첨자 부분입니다. 그런데 알파와 베타의 recurrence(수식...
@raulchen Any progress on this issue? or any alternative method for fractional gpu and several cpu worker mapping?
@mraj96 maybe discriminator will be too weak. but I'm gonna try it. training from scratch takes about 1 week in my environment. It sounds much doable for me.
@mraj96 finetuning or training hifigan takes really long time, so i gave up. i've finetuned vocoder with only generator, but ~300 epoch trainig showed nothing difference between using base vocoder....