Neural network-based singing voice synthesis library for research.
Neural network-based singing voice synthesis demo using kiritan_singing database (Japanese)
- Kiritan samples: https://soundcloud.com/r9y9/sets/dnn-based-singing-voice
- Python 3.6 or newer
- nnmnkwii: development version (master branch) is required
- pysinsy development version is required. Please have a look at the repostiory for installation.
- Pytorch >= 1.x
Note that packages listed above should be manually installed. After installing them, you can run:
python setup.py develop
to install the rest of dependencies.
- Core library: nnsvs/
- Command line programs: nnsvs/bin/ and its configurations nnsvs/bin/conf/
- Recipes: egs/
A recipe is a set of scripts and configuraitons that are used to reproduce experiments. All the steps used to conduct experiments are provided in a self-contained way. Please have a look at the egs directory if you want to build your singing voice systems.
As of Feb. 2020, NEUTRINO, a DNN-based singing voice synthesis tool, has started gaining its popularity in Japan. Because of the powerful DNN-based approach, users can create expressive and natural singing voices even without manual tuning which is typically required to achieve satisfactory quality using the existing tools.
While NEUTRINO is a great tool for creative purposes, it is not open-source software. In fact, there are only a few open-source toolkits to the best of our knowledge. To advance the singing voice synthesis research, we aim to provide a modern DNN-based singing voice synthesis tool for researchers and developers.
That being said, I was just curious to see if I can make a better one than NEUTRINO. We’ll see :)
- Hydra configurations are not easily overrided by users, without manually editing configs in nnsvs/bin/conf. See https://github.com/facebookresearch/hydra/issues/386 for details.