OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Audio & Voice AI

Hifi Gan

HiFi-GAN is an open-source, GAN-based speech synthesis model and Python implementation that generates raw WAV waveforms from mel-spectrograms, with additional documented WAV inference, training and fine-tuning workflows.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Hifi Gan — OpenViST project profile in Audio & Voice AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use HiFi-GAN when your workflow needs a model to turn generated mel-spectrograms into speech waveforms, or when you want to experiment with the repository’s pretrained weights, training or fine-tuning workflows.

Who it is for

  • Developers connecting a mel-spectrogram-producing text-to-speech model to waveform generation.
  • Researchers and practitioners who want to use pretrained speech-vocoder weights.
  • Users preparing speech data to train or fine-tune a generator.

Documented uses

  • Convert mel-spectrograms from systems such as Tacotron2 or Glow-TTS into WAV speech.
  • Run the documented inference workflow with WAV inputs and a generator checkpoint.
  • Train a generator with the provided configurations and LJ Speech dataset.
  • Fine-tune a pretrained generator using paired audio and mel-spectrogram data.
Before you choose

Requirements & limits

The implementation and pretrained models are provided in a public repository identified as open source under the MIT license.

  • End-to-end speech synthesis requires mel-spectrogram files; Tacotron2 and Glow-TTS are named as possible sources.
  • Training from scratch requires Python 3.6 or later, installed Python requirements, and LJ Speech files placed in LJSpeech-1.1/wavs.
  • The repository offers pretrained models associated with LJSpeech and VCTK, as well as a universal model described as a transfer-learning base.
  • The README describes WAV-file inference separately from mel-spectrogram inference.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Generate WAV audio from mel-spectrogram files

Use the end-to-end inference workflow to generate WAV output from mel-spectrograms.

Before you start

  • The HiFi-GAN repository and its Python requirements.
  • Generated mel-spectrogram files, such as those produced by Tacotron2 or Glow-TTS.
  • A generator checkpoint file.

Steps

  1. Create a test_mel_files directory and place the generated mel-spectrogram files in it.

  2. Run the command below.

    Documented command
    python inference_e2e.py --checkpoint_file [generator checkpoint file path]

  3. Find the generated WAV files in generated_files_from_mel; use --output_dir to change that directory.

Expected outcome

WAV files generated from the input mel-spectrograms.

Guide limitations

  • The repository documents this workflow for mel-spectrogram inputs; it does not describe text-to-speech text input in this workflow.
  • The guide follows repository instructions and does not verify execution or compatibility with a particular environment.
Documentation 1 ↗
Source-based practical guide

Fine-tune a pretrained model for a speech dataset

Follow the repository’s documented fine-tuning setup and command using matched audio and mel-spectrogram data.

Before you start

  • A pretrained HiFi-GAN model.
  • Audio files and matching mel-spectrograms in NumPy format, generated using Tacotron2 with teacher-forcing.
  • Python and the repository’s installed requirements.

Steps

  1. Generate a .npy mel-spectrogram for each audio file, using the same file name stem as the audio file.

  2. Create an ft_dataset folder and copy the generated mel-spectrogram files into it.

  3. Run the command below.

    Documented command
    python train.py --fine_tuning True --config config_v1.json

Expected outcome

Starts the documented fine-tuning workflow for adapting the generator to additional speech data.

Guide limitations

  • The README documents the command and input preparation but does not specify a guaranteed result or evaluation procedure.
  • The matching audio and mel-spectrogram naming requirement applies to the workflow.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
jik876/hifi-gan · Official homepage title · retained in its source language
How we review sourcesSuggest a correction