README_prev.md

metadata

language:
  - km
license: apache-2.0
tags:
  - automatic-speech-recognition
  - openslr
  - robust-speech-event
  - km
  - generated_from_trainer
model-index:
  - name: xls-r-300m-km
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: OpenSLR km
          type: openslr
          args: km
        metrics:
          - name: Test WER
            type: wer
            value: 29.26
          - name: Test CER
            type: cer
            value: 7.93

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the openslr dataset. It achieves the following results on the evaluation set:

• Loss: 0.3142
• Wer: 0.3512

Evaluation results on OpenSLR "evaluation" (self-split) (Running ./eval.py):

• WER: 0.2925882809468374
• CER: 0.0792776460744666

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• 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: 100
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.2049	4.93	400	4.5570	1.0
3.569	9.87	800	3.5415	1.0
3.483	14.81	1200	3.3956	1.0
2.1906	19.75	1600	1.1732	0.7897
1.7968	24.69	2000	0.7634	0.6678
1.615	29.62	2400	0.6182	0.5922
1.52	34.56	2800	0.5473	0.5479
1.4696	39.5	3200	0.5002	0.5130
1.4175	44.44	3600	0.4752	0.5021
1.3943	49.38	4000	0.4638	0.4944
Pause and Resume
1.3829	4.93	400	0.4290	0.4796
1.3156	9.87	800	0.3856	0.4474
1.2396	14.81	1200	0.3600	0.4307
1.1444	19.75	1600	0.3423	0.4179
1.0979	24.69	2000	0.3370	0.3884
1.0714	29.62	2400	0.3237	0.3710
1.0442	34.56	2800	0.3336	0.3683
1.0492	39.5	3200	0.3166	0.3527
1.0284	44.44	3600	0.3178	0.3566
1.0302	49.38	4000	0.3142	0.3512

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0