metadata

license: apache-2.0
tags:
  - generated_from_trainer
datasets:
  - google/fleurs
metrics:
  - wer
model-index:
  - name: xls-r-fleurs_nl-run3
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: audiofolder
          type: audiofolder
          config: default
          split: validation
          args: default
        metrics:
          - name: Wer
            type: wer
            value: 0.42659804983748645

xls-r-fleurs_nl-run3

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the FLEURS (nl) dataset. It achieves the following results:

Wer (Validation): 40.48%
Wer (Test): 40.89%

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 3
total_train_batch_size: 12
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
num_epochs: 30
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer (Train)
7.768	0.41	100	3.9649	1.0
3.2646	0.82	200	2.9551	1.0
2.9217	1.23	300	2.9128	1.0
2.9064	1.64	400	2.9067	1.0
2.6775	2.05	500	1.5774	0.9177
1.1026	2.47	600	0.8813	0.7216
0.6905	2.88	700	0.7287	0.6138
0.4936	3.29	800	0.6156	0.5439
0.3837	3.7	900	0.5608	0.4992
0.3176	4.11	1000	0.5326	0.4542
0.2391	4.52	1100	0.5221	0.4466
0.2426	4.93	1200	0.5127	0.4328
0.1882	5.34	1300	0.5311	0.4247
0.1718	5.75	1400	0.5523	0.4266

Framework versions

Transformers 4.28.0
Pytorch 2.0.1+cu117
Datasets 2.14.4
Tokenizers 0.13.3