---
language: en
license: mit
tags:
- speech
- audio
- vocoder
- hifigan
- tts
---

# HiFiGAN Finetuned Checkpoint

This is a finetuned HiFiGAN model checkpoint for vocoder.

## Model Details
- Base Model: HiFiGAN
- Training Type: Finetuned from HuggingFace checkpoint
- Sample Rate: 22050 Hz
- Checkpoint Date: 2025-03-30

## Files
- `generator.ckpt`: The model weights
- `hyperparams.yaml`: Model hyperparameters

## Usage
```python
from speechbrain.lobes.models.HifiGAN import HifiganGenerator
import torch
from hyperpyyaml import load_hyperpyyaml

# Load hyperparameters
with open("hyperparams.yaml") as f:
    hparams = load_hyperpyyaml(f)

# Initialize generator
generator = HifiganGenerator(
    in_channels=hparams["in_channels"],
    out_channels=hparams["out_channels"],
    resblock_type=hparams["resblock_type"],
    resblock_dilation_sizes=hparams["resblock_dilation_sizes"],
    resblock_kernel_sizes=hparams["resblock_kernel_sizes"],
    upsample_kernel_sizes=hparams["upsample_kernel_sizes"],
    upsample_initial_channel=hparams["upsample_initial_channel"],
    upsample_factors=hparams["upsample_factors"],
    inference_padding=hparams["inference_padding"],
    cond_channels=hparams["cond_channels"],
    conv_post_bias=hparams["conv_post_bias"],
)

# Load checkpoint
generator.load_state_dict(torch.load("generator.ckpt"))
generator.eval()
```