whisper-small-et-ERR2020

This model is a fine-tuned version of openai/whisper-small on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5321
• Wer: 22.8462

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: 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: 1000
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3387	0.1	1000	0.5383	33.8216
0.1393	0.2	2000	0.4897	27.7546
0.0982	0.3	3000	0.5477	26.7815
0.0912	1.02	4000	0.5195	24.8816
0.0811	1.12	5000	0.5373	25.9282
0.0649	1.22	6000	0.5422	23.7285
0.0618	1.32	7000	0.5504	23.5179
0.0558	2.03	8000	0.5321	22.8462
0.0452	2.13	9000	0.5543	23.5813
0.0462	2.23	10000	0.5424	22.8830

Framework versions

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