metadata
language:
- nl
license: apache-2.0
tags:
- whisper-event
- hf-asr-leaderboard
datasets:
- mozilla-foundation/common_voice_11_0
metrics:
- wer
model-index:
- name: Whisper Medium nl - GeoffVdr
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: mozilla-foundation/common_voice_11_0
type: mozilla-foundation/common_voice_11_0
config: nl
split: test
args: nl
metrics:
- name: Wer
type: wer
value: 7.514
co2_eq_emissions:
emissions: 2930
source: https://mlco2.github.io/impact/
training_type: fine-tuning
geographical_location: Ghent, Belgium
hardware_used: 1 v100 GPU
Whisper Medium nl - GeoffVdr
This model is a fine-tuned version of openai/whisper-medium on the Common Voice 11.0 dataset.
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
- Training: Mozilla CommonVoice 11 Dutch train+validation set
- Evaluation: Mozilla CommonVoice 11 Dutch test set
Training procedure
Training Hyperparameters
- learning_rate: 1e-5
- train_batch_size: 8
- eval_batch_size: 8
- gradient_accumulation_steps: 2
- lr_scheduler_warmup_steps: 500
- training_steps: 12000
Training Results
| Training Loss | Epoch | Step | WER |
|---|---|---|---|
| 0.1111 | 0.39 | 1000 | 9.89 |
| 0.0884 | 0.78 | 2000 | 9.26 |
| 0.0362 | 1.17 | 3000 | 8.64 |
| 0.0359 | 1.56 | 4000 | 8.60 |
| 0.0375 | 1.95 | 5000 | 8.24 |
| : | |||
| : | |||
| 0.0015 | 4.68 | 12000 | 7.51 |