|
--- |
|
license: cc0-1.0 |
|
--- |
|
|
|
# بسم الله |
|
|
|
1) Converted universal v1, v2, v3 22050Hz hifigan models from [here](https://github.com/jik876/hifi-gan) and [here](https://drive.google.com/drive/folders/1-eEYTB5Av9jNql0WGBlRoi-WH2J7bp5Y) |
|
|
|
2) Converted to onnx using [this](https://github.com/k2-fsa/icefall/blob/master/egs/ljspeech/TTS/matcha/export_onnx_hifigan.py) |
|
|
|
3) Added metadata as such using the above python script: |
|
{'model_type': 'hifigan', 'model_filename': 'hifigan_universal_22050hz_v1', 'sample_rate': 22050, 'version': 1, 'model_author': 'jik876', 'maintainer': 'k2-fsa', 'dataset': 'Universal', 'url1': 'https://github.com/jik876/hifi-gan', 'url2': ''} |
|
|
|
Note1: hifigan v2 is much smaller, faster and with lower memory need |
|
|
|
Note2: As LJ speech is a female single speaker dataset, hifigan based on it does not work well on men's voice. vctk models work well for both genders. |
|
|
|
Conclusion: Use vctk_v2! |
