legendary2910/Mnong-ASR-v1-enhanced

Whisper Medium Mnong

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

• Loss: 0.0471
• Wer: 7.2593

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
2.3012	0.1421	200	2.0190	142.3332
1.4133	0.2843	400	1.3463	99.4906
0.9797	0.4264	600	0.9503	80.8456
0.7402	0.5686	800	0.6821	62.0479
0.4908	0.7107	1000	0.4992	47.8349
0.3865	0.8529	1200	0.4090	42.6133
0.3031	0.9950	1400	0.3108	34.7937
0.203	1.1372	1600	0.2632	39.2511
0.1846	1.2793	1800	0.2209	28.3749
0.1313	1.4215	2000	0.1776	18.2119
0.0984	1.5636	2200	0.1525	18.8487
0.1009	1.7058	2400	0.1276	14.8242
0.0803	1.8479	2600	0.1034	12.1498
0.061	1.9900	2800	0.0910	11.7422
0.0327	2.1322	3000	0.0808	12.3535
0.026	2.2743	3200	0.0716	9.2970
0.024	2.4165	3400	0.0612	10.2649
0.0282	2.5586	3600	0.0552	8.0234
0.016	2.7008	3800	0.0488	7.8961
0.0237	2.8429	4000	0.0471	7.2593

Framework versions

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