在 100 小时的 LibriSpeech 数据上训练后,HyperConformer 和 Conformer 在 test-other 上的 WER 更高

batch_size, n_gpus, grad_accumulation_factor, lr_adam：16，1，1，0.001 conformer_8M： epoch: 10, lr: 2.58e-04, steps: 6450, optimizer: Adam - train loss: 2.51e+02 - valid loss: 1.96e+02, valid ACC: 2.02e-01, valid WER: 1.06e+02 epoch: 20, lr: 5.16e-04, steps: 12900, optimizer: Adam - train loss: 1.23e+02 - valid loss: 79.15, valid ACC: 6.35e-01, valid WER: 41.40 epoch: 30, lr: 7.74e-04, steps: 19350, optimizer: Adam - train loss: 66.70 - valid loss: 48.60, valid ACC: 7.97e-01, valid WER: 21.52 epoch: 40, lr: 9.84e-04, steps: 25800, optimizer: Adam - train loss: 53.02 - valid loss: 39.94, valid ACC: 8.34e-01, valid WER: 17.25 epoch: 50, lr: 8.80e-04, steps: 32250, optimizer: Adam - train loss: 41.96 - valid loss: 36.77, valid ACC: 8.55e-01, valid WER: 14.83 epoch: 60, lr: 8.04e-04, steps: 38700, optimizer: Adam - train loss: 36.05 - valid loss: 33.44, valid ACC: 8.65e-01, valid WER: 13.59 epoch: 70, lr: 7.44e-04, steps: 45150, optimizer: Adam - train loss: 32.15 - valid loss: 32.92, valid ACC: 8.68e-01, valid WER: 12.74 epoch: 80, lr: 6.96e-04, steps: 51600, optimizer: Adam - train loss: 29.23 - valid loss: 33.27, valid ACC: 8.71e-01, valid WER: 12.19 epoch: 90, lr: 6.56e-04, steps: 58050, optimizer: Adam - train loss: 27.22 - valid loss: 35.81, valid ACC: 8.71e-01, valid WER: 11.96 epoch: 100, lr: 6.23e-04, steps: 64500, optimizer: Adam - train loss: 25.85 - valid loss: 33.84, valid ACC: 8.74e-01, valid WER: 11.65 epoch: 110, lr: 5.94e-04, steps: 70950, optimizer: Adam - train loss: 24.51 - valid loss: 33.23, valid …