whisper-small-et

This model is a fine-tuned version of openai/whisper-small on the following datasets: Common Voice 11, VoxPopuli and FLEURS.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

Estonian data from Common Voice 11, VoxPopuli and FLEURS corpora as both training and validation sets. Tested on Common Voice 11 test set.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-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: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.1285	1.03	200	1.0640	53.4934
0.5163	2.05	400	0.6450	41.2428
0.2005	4.01	600	0.5600	36.6797
0.1188	5.03	800	0.5718	35.2847
0.0487	6.06	1000	0.5999	34.7500
0.0216	8.01	1200	0.6479	38.1906
0.016	9.04	1400	0.6655	39.5034
0.0085	10.06	1600	0.7027	33.9038
0.0079	12.02	1800	0.7207	39.5723
0.009	13.04	2000	0.7261	34.5973

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.12.1+rocm5.1.1
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2