search

begeekmyfriend / tacotron2

Forked from NVIDIA/tacotron2 and merged with Rayhane-mamah/Tacotron-2
BSD 3-Clause "New" or "Revised" License
81 stars 38 forks source link

how many data should i use to train this model？ #19

Open ArtemisZGL opened 4 years ago

ArtemisZGL commented 4 years ago

Thanks for your job first, I want to know that how many data should i use to train in your repo. I want to use the cmu arctic dataset for training english tts, but there is about one hour for a speaker, can it work to train in you repo？Because i use the nvidia repo to train but the result is bad. Also, the result is very different for different batch size. I also use some of the libritts data to train on nvdia repo, about 3hour for 10 speaker, but the result is quiet bad too. Do you have some idea about how to train on small dataset ？

begeekmyfriend commented 4 years ago

Multi-speaker is supported as well. For instance, you might collect 8 speakers and one hour for each corpus and record the directory in scripts/train_tacotron2.sh. Then the total amounts of data might help.

ArtemisZGL commented 4 years ago

@begeekmyfriend Thanks for your reply, but one speaker in libritts just 10-20 min, if i just use about 10 speaker of this, will it work?

begeekmyfriend commented 4 years ago

I have no idea about your circumstances but it seems too little for everyone's corpus. You might try yourself but the evaluation can not be warranted.

hassanShabbir1960 commented 4 years ago

Thankyou so much sir, for such an amazing work. @begeekmyfriend