Dataset question

[pavelxx1]

Hi, first thx for this great repo!) I have question

• For best result I need dataset with small duration length ie 10sec per wav or better when length above 1-2min for one speaker Thx!

[k2kobayashi]

If the feature vectors are extracted correctly, that's no problem. But I assumed every audio length is around 5s.

[pavelxx1]

Thx, 1) How many wav- files(dataset) need for good training result? 2) If I want convert speakerA to speakerB can I use small dataset of speakerA, ie 5-10 wav-files? Thx

[k2kobayashi]

1. vcc2020v1 and vcc2018v1 use about 1k utterances for training and it produces acceptable quality.
2. I didn't try such condition. I predict 10~30 utterances / speaker is minimum # of training utterances for 2 speaker modeling.