legendary2910/Mnong-ASR-v3

Whisper Base Mnong

This model is a fine-tuned version of openai/whisper-base on the MnongAudio-v2 dataset. It achieves the following results on the evaluation set:

• Loss: 0.7611
• Wer: 77.7127

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: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.7389	0.2915	200	2.6665	373.6628
2.2233	0.5831	400	2.2426	189.4549
1.8164	0.8746	600	1.8990	131.6353
1.5731	1.1662	800	1.6678	124.3760
1.4459	1.4577	1000	1.4828	95.8227
1.3009	1.7493	1200	1.3453	96.9689
1.0242	2.0408	1400	1.2264	89.9898
0.9227	2.3324	1600	1.1492	80.0815
0.9111	2.6239	1800	1.0539	83.2399
0.8831	2.9155	2000	0.9899	88.1814
0.5906	3.2070	2200	0.9452	84.5899
0.54	3.4985	2400	0.9017	79.6740
0.542	3.7901	2600	0.8713	72.2364
0.4606	4.0816	2800	0.8320	72.9241
0.4879	4.3732	3000	0.8172	75.4712
0.4033	4.6647	3200	0.7940	75.9552
0.4235	4.9563	3400	0.7737	73.2552
0.3638	5.2478	3600	0.7704	79.2155
0.383	5.5394	3800	0.7641	77.7382
0.3714	5.8309	4000	0.7611	77.7127

Framework versions

• Transformers 4.42.4
• Pytorch 2.4.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1