hubert-base-common-voice-vi-demo

This model is a fine-tuned version of facebook/hubert-base-ls960 on the common_voice_16_1 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5121
• Wer: 0.3678

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: 0.0001
• train_batch_size: 32
• 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: 1000
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.8731	1.14	500	3.5477	1.0
3.3329	2.28	1000	2.1928	1.0171
1.4603	3.42	1500	0.9074	0.6542
0.9413	4.57	2000	0.7490	0.5568
0.7664	5.71	2500	0.6418	0.5052
0.6719	6.85	3000	0.6240	0.4819
0.6261	7.99	3500	0.6048	0.4657
0.5771	9.13	4000	0.5555	0.4512
0.525	10.27	4500	0.5475	0.4392
0.4948	11.42	5000	0.5619	0.4261
0.4585	12.56	5500	0.5646	0.4280
0.4584	13.7	6000	0.5326	0.4168
0.4157	14.84	6500	0.5126	0.4038
0.4113	15.98	7000	0.5282	0.4004
0.3955	17.12	7500	0.5310	0.3959
0.3658	18.26	8000	0.4936	0.3886
0.3584	19.41	8500	0.5438	0.3895
0.3536	20.55	9000	0.5167	0.3860
0.3665	21.69	9500	0.5194	0.3842
0.3231	22.83	10000	0.5269	0.3866
0.315	23.97	10500	0.5219	0.3768
0.3191	25.11	11000	0.5054	0.3728
0.3264	26.26	11500	0.5068	0.3710
0.3014	27.4	12000	0.5009	0.3694
0.3055	28.54	12500	0.5066	0.3676
0.3098	29.68	13000	0.5121	0.3678

Framework versions

• Transformers 4.39.3
• Pytorch 2.2.2+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2