wav2vec2-base-1

This model is a fine-tuned version of facebook/wav2vec2-base on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.9254
• Wer: 0.3216

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
1.6597	2.2	1000	0.8904	0.5388
0.4751	4.41	2000	0.7009	0.3976
0.3307	6.61	3000	0.7068	0.3672
0.2574	8.81	4000	0.7320	0.3544
0.2096	11.01	5000	0.7803	0.3418
0.177	13.22	6000	0.7768	0.3423
0.1521	15.42	7000	0.8113	0.3375
0.1338	17.62	8000	0.8153	0.3325
0.1168	19.82	9000	0.8851	0.3306
0.104	22.03	10000	0.8811	0.3277
0.0916	24.23	11000	0.8722	0.3254
0.083	26.43	12000	0.9527	0.3265
0.0766	28.63	13000	0.9254	0.3216

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.13.3
• Tokenizers 0.10.3