ALM/whisper-da-small-augmented

Whisper Small Danish - Robust

This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 da dataset. It achieves the following results on the evaluation set:

• Loss: 0.7926
• Wer: 32.3251

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: 64
• eval_batch_size: 32
• 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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0232	15.15	1000	0.7538	35.5813
0.0061	30.3	2000	0.7933	34.3766
0.0016	45.45	3000	0.7993	33.5823
0.0003	60.61	4000	0.7986	31.6097
0.0002	75.76	5000	0.7901	32.1357

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.8.0
• Tokenizers 0.13.2