This model was trained by Gunnar Thor Örnólfsson in 2023
It uses a combination of the Talrómur 1 and 2 corpora, totalling 44 speakers (with 4 speakers being held back for evaluation.
All 8 voices in Talrómur 1 were used. 36 out of the 40 voices in Talrómur 2 were used.
The model was trained for 1.500.000 steps using Tomoki Hayashi's implementation of Style-MelGAN: https://github.com/kan-bayashi/ParallelWaveGAN/
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support