Pretrained model for NNSVS's ResF0NonAttentiveTacotron acoustic model type
Data used will not be shared on this repo.
Uses intunist's Japanese compatibility HED and dictionary included with their Google Colab training notebook.
- Language - Japanese_compatibility
- Acoustic model type - ResF0NonAttentiveTacotron
- use_mdn - False
- Vibrato mode - none
- force_fix_vuv - False
- filter_long_segments - False
- sample_rate - 44100
- d4c_threshold - 0.25
- acoustic_loss - mae
- pitch_reg_weight - 0.05
In intunist's Colab notebook, make and run a new cell like so: !wget {link to model release}
Make another cell to unzip the file: !unzip {path to model}
Expand the model's exp folder and copy+paste the path to the folder called ResF0NonAttentiveTacotron_pretrained_intunist_prototyping_notebook into the notebook's pretrained_expdir setting.
There is a total of 5 hours and 21 minutes of data used in this model.
| Singer | Minutes of data (without silence) |
|---|---|
| Ariel | 68 |
| アルパカ肉 | 6 |
| ATSUYA | 7 |
| 瓶詰め | 5 |
| ちかの | 4 |
| 九 | 4 |
| ふぇりす。 | 3 |
| Haruqa | 5 |
| アトリ科ヒワ属のゲン | 2 |
| いろは酢 | 3 |
| 高峯いと | 4 |
| いつり | 4 |
| 匿名男声 | 5 |
| jvs001 | 2 |
| jvs002 | 1 |
| jvs010 | 1 |
| jvs039 | 1 |
| jvs076 | 1 |
| かっぴりー | 3 |
| 草薙快速雷虎 | 5 |
| まいこ | 4 |
| MIZKI | 3 |
| おふとんP | 4 |
| 紺瀬ぷち | 4 |
| ささささ | 3 |
| 翠澤しのん | 6 |
| Suzu | 104 |
| Tetsu | 75 |
| とめあ | 3 |
| とろっぽ | 3 |
| Google Translate | 20 |
- jvs010, 039, and 076 labeled by alice
- jvs001, 002, and Google Translate labeled by Mildemelwe
- Ariel by めんるい
- Tetsu by たまご
- Suzu by ハイッヤー
- Other data is from https://github.com/oatsu-gh/enunu_kodoku_singing
- Trained on intunist's NNSVS training notebook