legendary2910/Mnong-ASR-v3-enhanced

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: 0.5864
• Wer: 73.4845

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
2.7525	0.1421	200	2.6537	416.5818
2.2459	0.2843	400	2.2237	158.5838
1.8682	0.4264	600	1.8896	237.7483
1.7212	0.5686	800	1.6295	110.0866
1.4164	0.7107	1000	1.4443	108.9913
1.2698	0.8529	1200	1.3000	91.3653
1.1479	0.9950	1400	1.1657	102.3688
1.0034	1.1372	1600	1.0799	84.6918
0.945	1.2793	1800	0.9844	85.7106
0.8249	1.4215	2000	0.8974	87.1880
0.726	1.5636	2200	0.8412	92.9699
0.7561	1.7058	2400	0.7859	80.8202
0.6884	1.8479	2600	0.7328	85.3031
0.6329	1.9900	2800	0.6872	80.9985
0.5129	2.1322	3000	0.6672	76.4901
0.5361	2.2743	3200	0.6369	78.2985
0.482	2.4165	3400	0.6178	75.9042
0.5211	2.5586	3600	0.6030	79.3938
0.4749	2.7008	3800	0.5905	76.0316
0.4648	2.8429	4000	0.5864	73.4845

Framework versions

• Transformers 4.43.4
• Pytorch 2.4.0+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1