fine-tuned-atc-speech-to-text-v36a

This model is a fine-tuned version of openai/whisper-medium.en on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.9518
• Wer Ortho: 19.4385
• Wer: 18.5857

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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• num_epochs: 15
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer
4.8487	0.9994	407	0.6427	47.5231	46.2577
1.7757	1.9988	814	0.5018	31.3702	30.4120
1.2488	2.9982	1221	0.7041	26.7013	25.6774
0.9809	4.0	1629	0.7811	23.8124	22.7863
0.7954	4.9994	2036	0.7675	22.2460	21.2854
0.6695	5.9988	2443	0.9535	21.5034	20.4795
0.5549	6.9982	2850	0.7755	20.0183	19.1397
0.4812	8.0	3258	1.0105	20.2319	19.4117
0.4149	8.9994	3665	1.0165	19.9878	19.2203
0.3701	9.9988	4072	1.0810	19.5301	18.7267
0.3321	10.9982	4479	0.9518	19.4385	18.5857
0.3107	12.0	4887	0.9795	19.8454	18.9181
0.2845	12.9994	5294	1.1289	19.7030	18.8375
0.2802	13.9988	5701	1.1235	19.5911	18.7166

Framework versions

• Transformers 4.45.2
• Pytorch 2.4.1+cu124
• Datasets 3.0.1
• Tokenizers 0.20.1