34j/neural-source-filter

Implements Neural Source Filter and NSF-HiFi-GAN vocoder architectures for neural audio synthesis, enabling high-quality speech generation through source-filter modeling. The package decouples pitch/spectral content from waveform generation, leveraging HiFi-GAN's discriminator-based training for improved audio quality. Targets PyTorch-based speech synthesis pipelines with pre-trained model support and integrates with standard audio processing workflows.

Available on PyPI.