This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the OPENSLR_SLR66 - NA dataset. It achieves the following results on the evaluation set:
- Loss: 0.3119
- Wer: 0.2613
Evaluation metrics
Metric | Split | Decode with LM | Value |
---|---|---|---|
WER | Train | No | 5.36 |
CER | Train | No | 1.11 |
WER | Test | No | 26.14 |
CER | Test | No | 4.93 |
WER | Train | Yes | 5.04 |
CER | Train | Yes | 1.07 |
WER | Test | Yes | 20.69 |
CER | Test | Yes | 3.986 |
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 4
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2000
- num_epochs: 150.0
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
2.9038 | 4.8 | 500 | 3.0125 | 1.0 |
1.3777 | 9.61 | 1000 | 0.8681 | 0.8753 |
1.1436 | 14.42 | 1500 | 0.6256 | 0.7961 |
1.0997 | 19.23 | 2000 | 0.5244 | 0.6875 |
1.0363 | 24.04 | 2500 | 0.4585 | 0.6276 |
0.7996 | 28.84 | 3000 | 0.4072 | 0.5295 |
0.825 | 33.65 | 3500 | 0.3590 | 0.5222 |
0.8018 | 38.46 | 4000 | 0.3678 | 0.4671 |
0.7545 | 43.27 | 4500 | 0.3474 | 0.3962 |
0.7375 | 48.08 | 5000 | 0.3224 | 0.3869 |
0.6198 | 52.88 | 5500 | 0.3233 | 0.3630 |
0.6608 | 57.69 | 6000 | 0.3029 | 0.3308 |
0.645 | 62.5 | 6500 | 0.3195 | 0.3722 |
0.5249 | 67.31 | 7000 | 0.3004 | 0.3202 |
0.4875 | 72.11 | 7500 | 0.2826 | 0.2992 |
0.5171 | 76.92 | 8000 | 0.2962 | 0.2976 |
0.4974 | 81.73 | 8500 | 0.2990 | 0.2933 |
0.4387 | 86.54 | 9000 | 0.2834 | 0.2755 |
0.4511 | 91.34 | 9500 | 0.2886 | 0.2787 |
0.4112 | 96.15 | 10000 | 0.3093 | 0.2976 |
0.4064 | 100.96 | 10500 | 0.3123 | 0.2863 |
0.4047 | 105.77 | 11000 | 0.2968 | 0.2719 |
0.3519 | 110.57 | 11500 | 0.3106 | 0.2832 |
0.3719 | 115.38 | 12000 | 0.3030 | 0.2737 |
0.3669 | 120.19 | 12500 | 0.2964 | 0.2714 |
0.3386 | 125.0 | 13000 | 0.3101 | 0.2714 |
0.3137 | 129.8 | 13500 | 0.3063 | 0.2710 |
0.3008 | 134.61 | 14000 | 0.3082 | 0.2617 |
0.301 | 139.42 | 14500 | 0.3121 | 0.2628 |
0.3291 | 144.23 | 15000 | 0.3105 | 0.2612 |
0.3133 | 149.04 | 15500 | 0.3114 | 0.2624 |
Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.11.0
- Downloads last month
- 35
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Dataset used to train chmanoj/xls-r-1B-te
Evaluation results
- Test WER on Open SLRself-reported20.624
- Test CER on Open SLRself-reported3.979
- Test WER (without LM) on Open SLRself-reported26.148
- Test CER (without LM) on Open SLRself-reported4.933