wav2vec2-base-timit-demo-colab

This model is a fine-tuned version of facebook/hubert-large-ls960-ft on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0280
• Wer: 0.0082

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
0.1152	1.42	500	0.0416	0.0159
0.0803	2.83	1000	0.0372	0.0144
0.0672	4.25	1500	0.0345	0.0119
0.0564	5.67	2000	0.0338	0.0106
0.0513	7.08	2500	0.0307	0.0100
0.0448	8.5	3000	0.0343	0.0098
0.0374	9.92	3500	0.0300	0.0084
0.0368	11.33	4000	0.0314	0.0086
0.0388	12.75	4500	0.0283	0.0089
0.0277	14.16	5000	0.0302	0.0089
0.0298	15.58	5500	0.0298	0.0089
0.0271	17.0	6000	0.0320	0.0098
0.024	18.41	6500	0.0286	0.0088
0.0236	19.83	7000	0.0284	0.0084
0.0238	21.25	7500	0.0290	0.0086
0.0227	22.66	8000	0.0284	0.0093
0.0198	24.08	8500	0.0280	0.0088
0.0225	25.5	9000	0.0281	0.0086
0.018	26.91	9500	0.0280	0.0082
0.0178	28.33	10000	0.0280	0.0082
0.0209	29.75	10500	0.0280	0.0082

Framework versions

• Transformers 4.15.0
• Pytorch 1.9.0+cu111
• Datasets 1.17.0
• Tokenizers 0.10.3