| model_name: "cnn_vad" | |
| # spec | |
| sample_rate: 8000 | |
| nfft: 512 | |
| win_size: 240 | |
| hop_size: 80 | |
| win_type: hann | |
| # model | |
| conv2d_block_param_list: | |
| - batch_norm: true | |
| in_channels: 1 | |
| out_channels: 8 | |
| kernel_size: 3 | |
| padding: "same" | |
| dilation: 3 | |
| activation: relu | |
| dropout: 0.1 | |
| - in_channels: 8 | |
| out_channels: 8 | |
| kernel_size: 5 | |
| padding: "same" | |
| dilation: 3 | |
| activation: relu | |
| dropout: 0.1 | |
| - in_channels: 8 | |
| out_channels: 8 | |
| kernel_size: 3 | |
| padding: "same" | |
| dilation: 2 | |
| activation: relu | |
| dropout: 0.1 | |
| encoder_output_size: 2056 | |
| # lsnr | |
| n_frame: 3 | |
| min_local_snr_db: -15 | |
| max_local_snr_db: 30 | |
| norm_tau: 1. | |
| # data | |
| min_snr_db: -10 | |
| max_snr_db: 20 | |
| # train | |
| lr: 0.001 | |
| lr_scheduler: "CosineAnnealingLR" | |
| lr_scheduler_kwargs: | |
| T_max: 250000 | |
| eta_min: 0.0001 | |
| max_epochs: 100 | |
| clip_grad_norm: 10.0 | |
| seed: 1234 | |
| num_workers: 4 | |
| batch_size: 128 | |
| eval_steps: 25000 | |