# jik876/hifi-gan architecture

> Architecture diagram of the GitHub repository jik876/hifi-gan, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-11-19). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/jik876/hifi-gan
- Repository: https://github.com/jik876/hifi-gan

## Overview

HiFi-GAN is a neural audio-generation project: it trains GAN vocoders and provides two inference workflows, converting either WAV input or precomputed mel spectrograms into generated WAV output. Training prepares audio/mel examples, optimizes a generator against period- and scale-based discriminators, and validates and checkpoints the model. Inference loads a configured checkpoint and runs the generator. The README also documents fine-tuning with externally generated mel files and V1/V2/V3 model configurations. Relationships below follow the excerpts and README; unsampled implementation details are not assumed.

Main entry points: [`train.py`](https://github.com/jik876/hifi-gan/blob/master/train.py) (Training Loop), [`inference.py`](https://github.com/jik876/hifi-gan/blob/master/inference.py) (WAV Inference), [`inference_e2e.py`](https://github.com/jik876/hifi-gan/blob/master/inference_e2e.py) (Mel Inference), [`models.py`](https://github.com/jik876/hifi-gan/blob/master/models.py) (Waveform Generator)

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_data["Audio Data"]
  node_wav_input["WAV Inputs"]
  node_mel_input["Mel Inputs"]
  node_manifests["Dataset Manifests<br/>[training.txt]"]
  node_mel_data["Mel Dataset<br/>[meldataset.py]"]
end

subgraph group_training["GAN Training"]
  node_config_v1["V1 Config<br/>[config_v1.json]"]
  node_config_v2["V2 Config<br/>[config_v2.json]"]
  node_config_v3["V3 Config<br/>[config_v3.json]"]
  node_trainer["Training Loop<br/>[train.py]"]
  node_training_utils["Training Utilities<br/>[utils.py]"]
end

subgraph group_models["Vocoder Models"]
  node_generator["Waveform Generator<br/>[models.py]"]
  node_period_disc["Period Discriminator<br/>[models.py]"]
  node_scale_disc["Scale Discriminator<br/>[models.py]"]
  node_gan_losses["GAN Losses<br/>[models.py]"]
  node_config_loader["Config Attributes<br/>[env.py]"]
end

subgraph group_inference["Audio Generation"]
  node_wav_inference["WAV Inference<br/>[inference.py]"]
  node_mel_inference["Mel Inference<br/>[inference_e2e.py]"]
  node_wav_output["Generated WAVs"]
end

node_user(("User"))

node_user -->|"starts training"| node_trainer
node_user -->|"runs inference"| node_wav_inference
node_user -->|"runs synthesis"| node_mel_inference
node_wav_input -->|"provides audio"| node_wav_inference
node_mel_input -->|"provides spectrograms"| node_mel_inference
node_manifests -->|"lists examples"| node_mel_data
node_trainer -->|"loads file lists"| node_manifests
node_trainer -->|"constructs datasets"| node_mel_data
node_trainer -->|"optimizes generator"| node_generator
node_trainer -->|"trains discriminator"| node_period_disc
node_trainer -->|"trains discriminator"| node_scale_disc
node_trainer -->|"computes losses"| node_gan_losses
node_trainer -->|"saves checkpoints"| node_training_utils
node_generator -->|"uses initialization"| node_training_utils
node_trainer -->|"selects configuration"| node_config_v1
node_trainer -->|"selects configuration"| node_config_v2
node_trainer -->|"selects configuration"| node_config_v3
node_wav_inference -->|"loads attributes"| node_config_loader
node_mel_inference -->|"loads attributes"| node_config_loader
node_trainer -->|"uses attributes"| node_config_loader
node_wav_inference -->|"computes mel features"| node_mel_data
node_wav_inference -->|"generates waveform"| node_generator
node_mel_inference -->|"generates waveform"| node_generator
node_wav_inference -->|"writes audio"| node_wav_output
node_mel_inference -->|"writes audio"| node_wav_output

click node_manifests "https://github.com/jik876/hifi-gan/blob/master/LJSpeech-1.1/training.txt"
click node_mel_data "https://github.com/jik876/hifi-gan/blob/master/meldataset.py"
click node_config_v1 "https://github.com/jik876/hifi-gan/blob/master/config_v1.json"
click node_config_v2 "https://github.com/jik876/hifi-gan/blob/master/config_v2.json"
click node_config_v3 "https://github.com/jik876/hifi-gan/blob/master/config_v3.json"
click node_trainer "https://github.com/jik876/hifi-gan/blob/master/train.py"
click node_training_utils "https://github.com/jik876/hifi-gan/blob/master/utils.py"
click node_generator "https://github.com/jik876/hifi-gan/blob/master/models.py"
click node_period_disc "https://github.com/jik876/hifi-gan/blob/master/models.py"
click node_scale_disc "https://github.com/jik876/hifi-gan/blob/master/models.py"
click node_gan_losses "https://github.com/jik876/hifi-gan/blob/master/models.py"
click node_config_loader "https://github.com/jik876/hifi-gan/blob/master/env.py"
click node_wav_inference "https://github.com/jik876/hifi-gan/blob/master/inference.py"
click node_mel_inference "https://github.com/jik876/hifi-gan/blob/master/inference_e2e.py"

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_wav_input,node_mel_input,node_manifests,node_mel_data,node_user toneBlue
class node_config_v1,node_config_v2,node_config_v3,node_trainer,node_training_utils toneAmber
class node_generator,node_period_disc,node_scale_disc,node_gan_losses,node_config_loader toneMint
class node_wav_inference,node_mel_inference,node_wav_output toneRose
```

## Components

### Audio Data

- **WAV Inputs**
- **Mel Inputs**
- [Dataset Manifests](https://github.com/jik876/hifi-gan/blob/master/LJSpeech-1.1/training.txt): `LJSpeech-1.1/training.txt`
- [Mel Dataset](https://github.com/jik876/hifi-gan/blob/master/meldataset.py): `meldataset.py`

### GAN Training

- [V1 Config](https://github.com/jik876/hifi-gan/blob/master/config_v1.json): `config_v1.json`
- [V2 Config](https://github.com/jik876/hifi-gan/blob/master/config_v2.json): `config_v2.json`
- [V3 Config](https://github.com/jik876/hifi-gan/blob/master/config_v3.json): `config_v3.json`
- [Training Loop](https://github.com/jik876/hifi-gan/blob/master/train.py): `train.py`
- [Training Utilities](https://github.com/jik876/hifi-gan/blob/master/utils.py): `utils.py`

### Vocoder Models

- [Waveform Generator](https://github.com/jik876/hifi-gan/blob/master/models.py): `models.py`
- [Period Discriminator](https://github.com/jik876/hifi-gan/blob/master/models.py): `models.py`
- [Scale Discriminator](https://github.com/jik876/hifi-gan/blob/master/models.py): `models.py`
- [GAN Losses](https://github.com/jik876/hifi-gan/blob/master/models.py): `models.py`
- [Config Attributes](https://github.com/jik876/hifi-gan/blob/master/env.py): `env.py`

### Audio Generation

- [WAV Inference](https://github.com/jik876/hifi-gan/blob/master/inference.py): `inference.py`
- [Mel Inference](https://github.com/jik876/hifi-gan/blob/master/inference_e2e.py): `inference_e2e.py`
- **Generated WAVs**

### Other components

- **User**

## Connections

- User → Training Loop: starts training (evidence: `README.md`)
- User → WAV Inference: runs inference (evidence: `README.md`)
- User → Mel Inference: runs synthesis (evidence: `README.md`)
- WAV Inputs → WAV Inference: provides audio (evidence: `README.md`)
- Mel Inputs → Mel Inference: provides spectrograms (evidence: `README.md`)
- Dataset Manifests → Mel Dataset: lists examples (evidence: `meldataset.py`)
- Training Loop → Dataset Manifests: loads file lists (evidence: `train.py`)
- Training Loop → Mel Dataset: constructs datasets (evidence: `train.py`)
- Training Loop → Waveform Generator: optimizes generator (evidence: `train.py`)
- Training Loop → Period Discriminator: trains discriminator (evidence: `train.py`)
- Training Loop → Scale Discriminator: trains discriminator (evidence: `train.py`)
- Training Loop → GAN Losses: computes losses (evidence: `train.py`)
- Training Loop → Training Utilities: saves checkpoints (evidence: `train.py`)
- Waveform Generator → Training Utilities: uses initialization (evidence: `models.py`)
- Training Loop → V1 Config: selects configuration (evidence: `README.md`)
- Training Loop → V2 Config: selects configuration (evidence: `README.md`)
- Training Loop → V3 Config: selects configuration (evidence: `README.md`)
- WAV Inference → Config Attributes: loads attributes (evidence: `inference.py`)
- Mel Inference → Config Attributes: loads attributes (evidence: `inference_e2e.py`)
- Training Loop → Config Attributes: uses attributes (evidence: `train.py`)
- WAV Inference → Mel Dataset: computes mel features (evidence: `inference.py`)
- WAV Inference → Waveform Generator: generates waveform (evidence: `inference.py`)
- Mel Inference → Waveform Generator: generates waveform (evidence: `inference_e2e.py`)
- WAV Inference → Generated WAVs: writes audio (evidence: `inference.py`)
- Mel Inference → Generated WAVs: writes audio (evidence: `inference_e2e.py`)
