[FastPitch] Other language

[Pawel-VRtechnology]

Hi everyone, I have a question about another language. Can I modify FastPitch to train other languages ​​like Polish ??

[alancucki]

Hi @Pawel-VRtechnology ,

That should be fairly straightforward. You're gonna need to adjust --text-cleaners, --symbol-set and either get ahold of a Polish pronunciation dictionary and supply it with --cmudict-path, or train on grapheme inputs with --p-arpabet=0.0.

Also, remember to set --pitch-mean and --pitch-std to your Polish speaker.

Our pre-trained WaveGlow checkpoint should work well for synthesis in Polish. For better results consider training HiFi-GAN.

[Pawel-VRtechnology]

@alancucki thanks for the answer, i will try your advice :)

[Moon-sung-woo]

Hi @alancucki I also train other languages like korean. But I'm going to train a multi-speaker. My question is : How to set --pitch-mean and --pitch-std? Do I have to get a mean for all the speakers' pitches?

[martinvk1]

@Moon-sung-woo I was facing this problem too. The current implementation doesn't support multiple mean/std values for each speakers, but clearly that is needed in the multispeaker setting. Calculate these values for each speaker, store them in a file and load them in the DataLoader. This worked well for me.