| license: mit | |
| language: | |
| - ru | |
| - en | |
| pipeline_tag: audio-to-audio | |
| # Summary | |
| Pretrained RingFormer v2 from [codename's RVC fork](https://github.com/codename0og/codename-rvc-fork-4) | |
| **Sample rate**: 40000 | |
| **Embedder**: Spin v2 | |
| **Mel Loss**: Experimental L1 + MR-STFT loss | |
| **F0**: RMVPE | |
| **Precision**: FP16 / FP32 mixed (BF16 proved to be very imprecise in ablation tests, achieving way worse loss values, I don't recommend using it for training any RVC models) | |
| **Speakers**: 225 | |
| **Dataset length**: 235:31:50, of which 37:26:53 is Russian, rest is English. | |
| **First checkpoints in the upcoming days, haven't finished training yet** |