| data: | |
| block_size: 512 | |
| cnhubertsoft_gate: 10 | |
| duration: 3 | |
| encoder: contentvec768l12 | |
| encoder_ckpt: pretrain/contentvec/checkpoint_best_legacy_500.pt | |
| encoder_hop_size: 320 | |
| encoder_out_channels: 768 | |
| encoder_sample_rate: 16000 | |
| extensions: | |
| - wav | |
| f0_extractor: rmvpe | |
| f0_max: 800 | |
| f0_min: 65 | |
| sampling_rate: 44100 | |
| train_path: data-v-f/train | |
| valid_path: data-v-f/val | |
| device: cuda | |
| env: | |
| expdir: exp/v-f-fp16-1 | |
| gpu_id: 1 | |
| infer: | |
| method: pndm | |
| speedup: 10 | |
| model: | |
| k_step_max: 200 | |
| n_chans: 512 | |
| n_layers: 20 | |
| n_spk: 1 | |
| pcmer_norm: true | |
| type: DiffusionNew | |
| use_pitch_aug: true | |
| train: | |
| amp_dtype: fp16 | |
| batch_size: 40 | |
| cache_all_data: true | |
| cache_device: cuda | |
| cache_fp16: true | |
| decay_step: 50000 | |
| epochs: 10000 | |
| gamma: 0.5 | |
| interval_force_save: 2000 | |
| interval_log: 1 | |
| interval_val: 1000 | |
| lambda_ddsp: 1 | |
| lr: 0.00015 | |
| num_workers: 0 | |
| save_opt: false | |
| weight_decay: 0 | |
| vocoder: | |
| ckpt: pretrain/nsf_hifigan/model | |
| type: nsf-hifigan | |