wav2vec2-large_phoneme-timit_english_timit-4k_001

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

• Loss: 0.4952
• Per: 0.1134

Model description

The wav2vec 2.0 large model is pre-trained on 960 hours of the LibriSpeech dataset.

• 24 Transformer blocks (Each block: 1024 dimensions & 16 attention heads)

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: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 5000
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Per
4.5458	3.46	1000	0.9087	0.2354
0.7877	6.92	2000	0.4441	0.1506
0.5125	10.38	3000	0.4241	0.1451
0.4485	13.84	4000	0.4244	0.1461
0.4193	17.3	5000	0.4618	0.1510
0.3899	20.76	6000	0.4700	0.1469
0.3244	24.22	7000	0.4496	0.1438
0.2717	27.68	8000	0.4988	0.1455
0.2222	31.14	9000	0.5182	0.1414
0.1872	34.6	10000	0.5320	0.1411

Framework versions

• Transformers 4.28.1
• Pytorch 2.0.1
• Datasets 2.18.0
• Tokenizers 0.13.3