scarlet coral
This is a test model based on a newly pre-trained model built on HiFi-GAN. The primary goal of this pre-trained model is to serve as a foundation specifically optimized for LIVE Voice Changer users.
Below are the key focus areas of this experimental pre-trained model:
A model capable of reproducing non-verbal human sounds, not just voice.
Removal of pitch constraints to enable the generation of high-pitched female laughter, coughing, and screaming.
Support for microphone artifacts, such as "pop" sounds caused by breath or touching the mic.
New embedding model training (fine-tuned on HuBERT {in Experimental Version, it will only Support Korean and Japanese}).
This version of the "NELL" model is an extension of the original, trained with additional samples including screams, choking, gagging, surprised reactions, coughing, and laughter. It is designed to test the model’s ability to handle whispers and extremely high-pitched vocalizations.
If you're using this model with tools like W-Okada, any feedback would be greatly appreciated and will help improve future versions.
Model Link - https://huggingface.co/SeoulStreamingStation/RVC_Voice_Models/resolve/main/Voice_Nell_Xe4_weightsgg.zip?download=true
Information
Train -
939 Epochs
Batch Size per GPU 16 x2 (Total 32)
LR : 5e-5
Pretrained model : KLM VoiceConvert Exp.
HIFIGAN / 32fps
Emb. model : ContentVec
Data -
216 Mins of Speech
23 Mins of Sing
10 Mins of SFX
10 Mins of Unclean Data
rapid sierra
winged slate
paper aspen
weary osprey