Closed chmh19961030 closed 5 years ago
I think that I have found some ideas on this issue. There are 2 versions of Tacotron-2 on my computer. the latest one was published a few days ago by @Rayhane-mamah and the older one was published a few months ago. the Tacotron2(Frontend and backend) is trained on the older Tacotron-2. When I finished the training of frontend and backend, the latest one is published. So I copied the logs-Tacotron-2 folder to the latest one and started the training of Wavenet.......althongh error did not occured during the running of frontend and backend, the output is wrong. When I run the same model in the older code, with text"Scientists at the CERN laboratory say they have discovered a new particle.", npy file output by Tacotron2 only contains a 416×80 array. I will continue doing research on this and report my idea:) thank you very much!
Hello, I'm facing the same issue, did you get any way around that?
thank you
Hi, everyone. Firstly, I'd like to thank you for your excellent Tacotron-2 code! @Rayhane-mamah I'm a rookie in TTS, and I'm trying to run this Tacotron-2 model on my server. But I came up with a problem. I was training the end-to-end TTS with LJSpeech. When I listen to the wavs in eval-dir/wavs after 60k wavenet training step, I found it was good. So I tried to synthesis wav from TTS. In the first time, I tried 'Hello.' But wavenet told me that it need generate about 80000 samples. The sample rate was 22050, which means that 'Hello' takes about 4 seconds! When it finished, I found that the wav is horrible..... I can hear something like hello, but other part of this audio is completely mess. In the second time, I tried 'Scientists at the CERN laboratory say they have discovered a new particle.' And Wavenet told me that it need 2560000 samples! WOW! So, I checked the npy file created by Tacotron2, and I found that it is right. The npy file contains a 10000×80 array.I'm not sure whether the output is right. So the question is: why did Wavenet try to generate so many samples? Would you please be so kind to help me on this problem? @Rayhane-mamah
I'd like to provide my hparams here. I have changed some of them because I'd like to comparte Tacotron2 with other mel-spectrogram generator...
tacotron_num_gpus = 1, #Determines the number of gpus in use for Tacotron training. wavenet_num_gpus = 1, #Determines the number of gpus in use for WaveNet training. split_on_cpu = True, #Determines whether to split data on CPU or on first GPU. This is automatically True when more than 1 GPU is used.
(Recommend: False on slow CPUs/Disks, True otherwise for small speed boost)