whisper-atcosim3

This model is a fine-tuned version of openai/whisper-medium on the atco2_atcosim dataset. It achieves the following results on the evaluation set:

• Loss: 0.0476
• Wer: 0.0198

Model description

This model is a special ASR model, derived doing fine-tuning of OpenAI Whisper model on ATC conversations. The base model is: OpenAI Whisper Medium

Intended uses & limitations

More information needed

Training and evaluation data

The model has been trained on the atco2_atcosim dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 150
• training_steps: 600
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.8218	0.2	50	0.3785	0.1451
0.1429	0.39	100	0.1213	0.0714
0.1155	0.59	150	0.0807	0.0517
0.0764	0.79	200	0.0652	0.0272
0.0724	0.98	250	0.0607	0.0393
0.0357	1.18	300	0.0569	0.0242
0.03	1.38	350	0.0553	0.0243
0.0325	1.57	400	0.0556	0.0228
0.03	1.77	450	0.0501	0.0242
0.0232	1.97	500	0.0485	0.0205
0.0143	2.16	550	0.0480	0.0194
0.0105	2.36	600	0.0476	0.0198

Framework versions

• Transformers 4.29.0
• Pytorch 2.0.0+cu117
• Datasets 2.12.0
• Tokenizers 0.11.0