| This model is based on whisper large v3 and trained on the dataset of jlvdoorn/atco2-asr-atcosim | |
| It starts as an experimental model and seems to be abit overfit to air traffic vocabs but with better capturing of some of the radio words | |
| ``` | |
| Training Metric: | |
| eval/loss:0.06807401776313782 | |
| eval/runtime:1,922.2684 | |
| eval/samples_per_second:1.054 | |
| eval/steps_per_second:0.132 | |
| eval/wer:2.3947324274450597 | |
| train/epoch:7.1146245059288535 | |
| train/global_step:3,600 | |
| train/grad_norm:0.05588585510849953 | |
| train/learning_rate:0.00000406779661016949 | |
| train/loss:0.001 | |
| ``` | |
| hyperparameter: | |
| ``` | |
| --model_name openai/whisper-large-v3 \ | |
| --train_batch_size 16 \ | |
| --eval_batch_size 8 \ | |
| --bf16 false \ | |
| --learning_rate 1e-5 \ | |
| --warmup_steps 100 \ | |
| --max_steps 6000 \ | |
| --save_steps 200 \ | |
| --eval_steps 200 \ | |
| --gradient_checkpointing true \ | |
| --output_dir ./whisper_atc_20241219 | |
| ``` |