Universal vocoders for various synthesis toolkits and architectures
Trained on 350hr+ of audio data from a wide variety of voice types and languages.
Vocoders are provided in two flavors, common and community.
The "common" vocoder utilizes only data that {the team} controls and data with open licenses/lack thereof.
This vocoder utilizes the base ~350hr of data with no additions.
The "community" vocoder utilizes data submitted by users for inclusion in the vocoder dataset. This data may have it's own licensing that MUST be followed on top of the license for the vocoder itself. Due to potential complications caused by this it may be desirable for some to use the "common" vocoder instead.
It should ve noted that due to the make-up of the community that the community vocoder could develop a significant bias for higher pitch male voices. Some members of our team have bestowed this vocoder with the nickname "the twinkoder".
The "common" vocoder is trained on a singing range of F1 ~ C7.
For pretraining purposes it's strongly advised to use the "common" vocoders as a better base.
Examples: team_sifigan-world_common / team_sifigan-world_community