metadata
language:
- en
license: apache-2.0
tags:
- hf-asr-leaderboard
- generated_from_trainer
datasets:
- atcosim_corpus
metrics:
- wer
model-index:
- name: atcosim_corpus
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: atcosim_corpus
type: atcosim_corpus
args: 'split: test'
metrics:
- name: Wer
type: wer
value: 2.490946029502694
atcosim_corpus
This model is a fine-tuned version of openai/whisper-tiny on the atcosim_corpus dataset. It achieves the following results on the evaluation set:
- Loss: 0.0623
- Wer: 2.4909
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: 1e-05
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.0281 | 2.1 | 1000 | 0.0716 | 4.1957 |
| 0.0051 | 4.19 | 2000 | 0.0650 | 2.7162 |
| 0.0009 | 6.29 | 3000 | 0.0624 | 2.4733 |
| 0.0005 | 8.39 | 4000 | 0.0623 | 2.4909 |
Framework versions
- Transformers 4.25.1
- Pytorch 1.13.0+cu116
- Datasets 2.6.1
- Tokenizers 0.13.2