legendary2910/Mnong-ASR-v2-enhanced

Whisper Small Mnong

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

• Loss: 0.1380
• Wer: 29.9287

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
3.2102	0.1421	200	3.0988	153.0565
1.7796	0.2843	400	1.7393	146.0774
1.3216	0.4264	600	1.3372	109.1187
1.0883	0.5686	800	1.0383	101.5028
0.8187	0.7107	1000	0.8161	63.4997
0.652	0.8529	1200	0.6821	66.2252
0.5411	0.9950	1400	0.5551	58.2272
0.4082	1.1372	1600	0.4738	58.5074
0.359	1.2793	1800	0.4075	45.1859
0.2761	1.4215	2000	0.3466	43.9379
0.212	1.5636	2200	0.3002	42.0785
0.2192	1.7058	2400	0.2642	36.0927
0.1932	1.8479	2600	0.2269	39.3785
0.1541	1.9900	2800	0.2013	30.5400
0.0944	2.1322	3000	0.1894	36.6021
0.0848	2.2743	3200	0.1682	29.4447
0.0811	2.4165	3400	0.1565	28.0183
0.0899	2.5586	3600	0.1481	31.0749
0.0749	2.7008	3800	0.1409	25.6240
0.0737	2.8429	4000	0.1380	29.9287

Framework versions

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