NNSVS

What NNSVS is

NNSVS is an open-source library for creating neural network-based singing voice synthesis systems, designed for research. It lets users build voicebanks from their own datasets and provides reproducible research recipes modeled on Kaldi and ESPnet. The recipes cover preprocessing, training, and synthesis and can be adapted to custom data and models. With a packed model, users can generate singing voice from MusicXML or UST files; WAV is the listed export format. The documented synthesis pipeline includes time-lag, duration, acoustic, post-filter, and vocoder modules. Users can implement custom PyTorch models through the NNSVS BaseModel interface. Optuna supports hyperparameter optimization, and MLflow tracks and visualizes search results. The documentation describes singing voice systems created with NNSVS for eight or more languages. It runs on Linux, Mac OS X, and Windows, with Linux recommended for development. Installation requires Python 3.7 or later and a C/C++ compiler; GPU/CUDA is strongly recommended for performance. Google Colab is an option for users without a GPU, but is not recommended for hyperparameter optimization. Acoustic model training can take several hours or a whole day, depending on configuration. The project is MIT-licensed, and commercial use is allowed.