search

jik876 / hifi-gan

HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis
MIT License
1.92k stars 506 forks source link

The effect of audio Split or not split? #69

Open WhiteFu opened 3 years ago

WhiteFu commented 3 years ago

Does audio segmentation(split) in data processing affect the model performance? Is it just for running the model to be able to support more batch size?

jik876 commented 3 years ago

Thanks for your interest. We found that there was a trade-off between segment size and audio quality in the early stage of our work, and we chose the segment size to train the model with small memory and fast iterations while achieving good quality.