whisper-large-et-ERR2020-v2

This model is a fine-tuned version of openai/whisper-large-v2 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2915
• Wer: 13.8640

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: 2
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• 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
0.2158	0.1	1000	0.3205	23.8154
0.0897	0.2	2000	0.2961	18.3340
0.0785	0.3	3000	0.2839	17.5230
0.0653	0.4	4000	0.2847	17.8752
0.0541	0.5	5000	0.2906	15.2645
0.0566	0.6	6000	0.2845	15.2081
0.051	0.7	7000	0.2888	14.4668
0.049	1.03	8000	0.2927	15.3130
0.044	1.13	9000	0.2915	13.8640
0.0379	1.23	10000	0.2913	16.5773

Framework versions

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