Speech Synthesis: HifiGAN Model Card

Model overview

HifiGAN is a neural vocoder model for text-to-speech applications. It is intended as the second part of a two-stage speech synthesis pipeline, with a mel-spectrogram generator such as FastPitch as the first stage.

Model architecture

HifiGAN is a neural vocoder based on a generative adversarial network framework, During training, the model uses a powerful discriminator consisting of small sub-discriminators, each one focusing on specific periodic parts of a raw waveform. The generator is very fast and has a small footprint, while producing high quality speech.

Training

Dataset

This model is trained on a proprietary dataset sampled at 44100Hz, and can be used to generate English voices with an American accent. This model supports 1 male voice and 1 female voice that comes from the Riva Fastpitch model.

How to use this model

HifiGAN is intended to be used as the second part of a two stage speech synthesis pipeline. HifiGAN takes a mel spectrogram and returns audio.

Input: Mel spectrogram of shape (batch x mel_channels x time)

Output: Audio of shape (batch x time)

Limitations

N/A

References

HifiGAN paper: https://arxiv.org/abs/2010.05646

License

By downloading and using the models and resources packaged with TLT Conversational AI, you would be accepting the terms of the Riva license.

Ethical AI

NVIDIA’s platforms and application frameworks enable developers to build a wide array of AI applications. Consider potential algorithmic bias when choosing or creating the models being deployed. Work with the model’s developer to ensure that it meets the requirements for the relevant industry and use case; that the necessary instruction and documentation are provided to understand error rates, confidence intervals, and results; and that the model is being used under the conditions and in the manner intended.