Aspik101/whisper-tiny-pl

Whisper Tiny Pl

This model is a fine-tuned version of openai/whisper-tiny on the Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.6919
• Wer: 39.4696

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: 16
• 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
0.3647	1.05	500	0.6999	43.0430
0.2752	3.04	1000	0.6002	39.3275
0.2513	5.04	1500	0.5911	37.8643
0.1661	7.04	2000	0.6381	37.6887
0.154	9.03	2500	0.6157	37.9362
0.1126	11.03	3000	0.6319	38.6402
0.0763	13.02	3500	0.6539	38.9730
0.0729	15.02	4000	0.6583	38.9362
0.0778	17.02	4500	0.6653	39.6769
0.0698	19.01	5000	0.6919	39.4696

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1
• Tokenizers 0.13.2