wav2vec2-base-finetuned-spgispeech-dev

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

• Loss: 0.2897
• Wer: 0.1508

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: 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: 1000
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.8285	2.22	1500	0.3361	0.2754
0.2582	4.44	3000	0.2643	0.2205
0.1697	6.66	4500	0.2467	0.2006
0.1314	8.88	6000	0.2711	0.1927
0.1084	11.09	7500	0.2521	0.1872
0.0922	13.31	9000	0.2588	0.1827
0.0818	15.53	10500	0.2572	0.1783
0.0712	17.75	12000	0.2720	0.1766
0.067	19.97	13500	0.2873	0.1751
0.0594	22.19	15000	0.2753	0.1704
0.0546	24.41	16500	0.2794	0.1694
0.0505	26.63	18000	0.2811	0.1665
0.0467	28.85	19500	0.2906	0.1657
0.0417	31.07	21000	0.3043	0.1661
0.0395	33.28	22500	0.3068	0.1627
0.0368	35.5	24000	0.3096	0.1617
0.0334	37.72	25500	0.3036	0.1581
0.0322	39.94	27000	0.2819	0.1564
0.0286	42.16	28500	0.2936	0.1544
0.0279	44.38	30000	0.2914	0.1534
0.0264	46.6	31500	0.2957	0.1519
0.0241	48.82	33000	0.2897	0.1508

Framework versions

• Transformers 4.17.0
• Pytorch 1.12.1+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1