metadata
license: cc-by-nd-4.0
tags:
- whisper-event
- generated_from_trainer
datasets:
- mozilla-foundation/common_voice_11_0
metrics:
- wer
model-index:
- name: odia_whisper_small_v3.0
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: mozilla-foundation/common_voice_11_0 or
type: mozilla-foundation/common_voice_11_0
config: or
split: test
args: or
metrics:
- name: Wer
type: wer
value: 13.8614
language:
- or
Ranjit/whisper_small_35k_or
This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 and other dataset. It achieves the following results on the evaluation set:
- Loss: 0.2584
- Wer: 13.8614
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-06
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 200
- training_steps: 5000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.0022 | 5.59 | 1000 | 0.2405 | 13.9995 |
| 0.0005 | 11.17 | 2000 | 0.2584 | 13.8614 |
| 0.0002 | 16.76 | 3000 | 0.2683 | 16.5598 |
| 0.0 | 22.35 | 4000 | 0.2907 | 15.0380 |
| 0.0 | 27.93 | 5000 | 0.3085 | 14.2035 |