cdli-whisper-en-ug-sunbird-encoder-a40

This model is a fine-tuned version of Sunbird/asr-whisper-large-v3-salt on an unknown dataset. It achieves the following results on the evaluation set:

Loss: 0.9443
Wer: 0.3339
Cer: 0.2373

Test Results

epoch = 7.7177
test_cer = 0.1415
test_loss = 0.6851
test_runtime = 0:10:08.37
test_samples_per_second = 1.665
test_steps_per_second = 0.418
test_wer = 0.2246

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: 2e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 16
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 150
training_steps: 2500

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.7206	0.7734	250	0.9708	0.3777	0.2721
0.5798	1.5445	500	0.9283	0.3481	0.2501
0.5444	2.3155	750	0.9325	0.3467	0.2508
0.5114	3.0866	1000	0.9377	0.3461	0.2485
0.4933	3.8600	1250	0.9307	0.3341	0.2398
0.4684	4.6311	1500	0.9362	0.3433	0.2463
0.4699	5.4022	1750	0.9397	0.3278	0.2327
0.4224	6.1732	2000	0.9451	0.3331	0.2363
0.4961	6.9466	2250	0.9438	0.3339	0.2372
0.4316	7.7177	2500	0.9443	0.3339	0.2373