Takenori Yoshimura
Takenori Yoshimura
When I ran a TTS recipe (e.g, egs2/siwis/tts1), I got some following messages. ``` [hoge] 2023-11-27 18:08:30,388 (trainer:753) INFO: 1epoch:train:1-25batch: iter_time=0.149, forward_time=1.375, l1_loss=1.740, mse_loss=2.692, bce_loss=0.198, attn_loss=0.010, loss=4.639, backward_time=1.206, grad_norm=3.849, clip=100.000,...
The new version of scipy (1.14) removed `csr_matrix.H`. https://docs.scipy.org/doc/scipy-1.13.0/reference/generated/scipy.sparse.csr_matrix.H.html Could you rewrite the following line? https://github.com/librosa/librosa/blob/cd66e10edf5fad2e95267b8289463848d54c39b7/librosa/core/constantq.py#L711
This improves the stability of the Taylor version of the MLSA filter. The order of the Taylor expansion can be reduced.
I fixed some bugs in MulticueF0v14 and modified it to support frame shifts other than 1 ms. I am not entirely sure if the modification is correct, but it appears...