Model
NVIDIA Deep Learning Examples
Jasper checkpoint (PyTorch, AMP, LibriSpeech)Jasper PyTorch checkpoint trained on LibriSpeech (test-other 9.66% WER)
19.08.0
Copied!
2 Versions
02/10/2021 12:37 AM UTC1.24 GBAccuracy: 0440 EpochsBatch Size: 512GPU: A100 Copied!
02/10/2021 12:37 AM UTC1.24 GBAccuracy: 0440 EpochsBatch Size: 512GPU: A100
Copied!
architecture
| Key | Value |
|---|---|
| type | jasper10x5dr_speedp-online_speca: SpecAugment (2x20, 3x75); Speed perturbation U(0.85, 1.15) |
performance
| Key | Value |
|---|---|
| dev_clean_wer | 3.20 |
| dev_other_wer | 9.74 |
| test_clean_wer | 3.38 |
| test_other_wer | 9.66 |
training
| Key | Value |
|---|---|
| batch_size | 512 |
| dataset | LibriSpeech |
| epochs | 440 |
| training_precision | AMP |
19.08.0Selected08/21/2020 1:36 AM UTC2.48 GBAccuracy: 0400 EpochsBatch Size: 0GPU: V100 Copied!
19.08.0Selected
08/21/2020 1:36 AM UTC2.48 GBAccuracy: 0400 EpochsBatch Size: 0GPU: V100
Copied!
architecture
| Key | Value |
|---|---|
| dense_residual_connnections | True |
| layers | 50 conv layers - 10 blocks - each 5 subblocks |
| model_configs | jasper10x5dr_sp_offline_specaugment.toml |
performance
| Key | Value |
|---|---|
| WER (on Librispeech dev-clean) | <3.64% |
training
| Key | Value |
|---|---|
| bs | 512 |
| cudnn | benchmark |
| epochs | 400 |
| gpu | V100 |
| gpus | 8 |
| grad_acc_steps | 1 |
| lr | 0.015 |
| lr_decay | quadratic |
| optim | novograd |
| spec_aug | True |
| speed_perturb | 1.0;0.9;1.1 |
| wd | 1e-3 |