whisper-large-v3-verbatim-1

This model is a fine-tuned version of openai/whisper-large-v3 on the JacobLinCool/ami-disfluent dataset. It achieves the following results on the evaluation set:

• Loss: 0.1300
• Wer: 32.3225
• Cer: 45.5147
• Decode Runtime: 141.5643
• Wer Runtime: 0.1227
• Cer Runtime: 0.2049

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: 5e-05
• train_batch_size: 4
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 1000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Decode Runtime	Wer Runtime	Cer Runtime
No log	0	0	1.8283	63.2783	251.8035	164.5307	0.1838	0.3386
0.2617	0.1	100	0.2189	49.6995	178.3721	161.1098	0.1397	0.4071
0.1291	0.2	200	0.1452	50.3383	95.5275	143.0863	0.1342	0.2932
0.1418	0.3	300	0.1387	29.9186	74.6491	150.1053	0.0780	0.1514
0.1273	1.088	400	0.1372	30.8218	91.1134	166.0178	0.1252	0.2728
0.1139	1.188	500	0.1335	29.9117	101.9003	144.2796	0.1318	0.2934
0.1663	1.288	600	0.1306	31.8418	83.0183	149.9060	0.0826	0.1679
0.1275	2.076	700	0.1311	24.9665	29.6191	143.2151	0.0781	0.1135
0.1077	2.176	800	0.1304	25.9109	36.6217	143.4620	0.0770	0.1227
0.1711	2.276	900	0.1298	35.1729	45.0300	145.3294	0.0786	0.1310
0.0994	3.064	1000	0.1300	32.3225	45.5147	141.5643	0.1227	0.2049

Framework versions

• PEFT 0.14.0
• Transformers 4.48.0
• Pytorch 2.4.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0