metadata
language:
- multilingual
license: apache-2.0
base_model: openai/whisper-medium
tags:
- hf-asr-leaderboard
- generated_from_trainer
datasets:
- abiyo27/BibleTTS_Ewe-Bible
metrics:
- wer
model-index:
- name: Whisper_Small_Ewe
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: BibleTTS
type: abiyo27/BibleTTS_Ewe-Bible
config: default
split: None
args: 'config: ewe, split: train'
metrics:
- name: Wer
type: wer
value: 10.094952523738131
Whisper_Small_Ewe
This model is a fine-tuned version of openai/whisper-medium on the BibleTTS dataset. It achieves the following results on the evaluation set:
- Loss: 0.1021
- Wer: 10.0950
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: 1
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 14000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.2196 | 0.1802 | 4000 | 0.1780 | 19.3903 |
| 0.1587 | 0.3605 | 8000 | 0.1375 | 13.4933 |
| 0.1162 | 0.5407 | 12000 | 0.1021 | 10.0950 |
Framework versions
- Transformers 4.41.2
- Pytorch 2.3.1+cu121
- Datasets 2.19.2
- Tokenizers 0.19.1