search

shivammehta25 / Matcha-TTS

[ICASSP 2024] 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching
https://shivammehta25.github.io/Matcha-TTS/
MIT License
716 stars 87 forks source link

What parameters could you suggest to tweak to improve model perfomance? #52

Closed Subuday closed 8 months ago

Subuday commented 8 months ago

Hi, I try to train your model on my personal datasets.

Are there any parameters you could suggest to tweak during training on my dataset? Or default params are enough for good quality results?

Are there any specific requirements to a dataset? I have the dataset of similar size like LJSpeech, 22050Hz, mono, similar duration distribution.

shivammehta25 commented 8 months ago

Hey!

Thank you for your interest in 🍵 Matcha-TTS.

Are there any parameters you could suggest to tweak during training on my dataset? Or default params are enough for good quality results?

It should work just fine with the default parameters, but I've noticed that you can improve quality by adding more parameters to the decoder. https://github.com/shivammehta25/Matcha-TTS/blob/256adc55d3219053d2d086db3f9bd9a4bde96fb1/configs/model/decoder/default.yaml#L1-L3

Something like channel 512, 512 and 128 attention dim.

Are there any specific requirements to a dataset? I have the dataset of similar size like LJSpeech, 22050Hz, mono, similar duration distribution.

This seems to be good enough.

Subuday commented 8 months ago

Thanks for your reply!