wav2vec2-large-xlsr-53-demo-colab

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 6.7860
• Wer: 1.1067

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.2273	44.42	400	3.3544	1.0
0.9228	88.84	800	4.7054	1.1601
0.1423	133.32	1200	5.9489	1.1578
0.0751	177.74	1600	5.5939	1.1717
0.0554	222.21	2000	6.1230	1.1717
0.0356	266.63	2400	6.2845	1.1613
0.0288	311.11	2800	6.6109	1.2100
0.0223	355.53	3200	6.5605	1.1299
0.0197	399.95	3600	7.1242	1.1682
0.0171	444.42	4000	7.2452	1.1578
0.0149	488.84	4400	7.4048	1.0684
0.0118	533.32	4800	6.6227	1.1172
0.011	577.74	5200	6.7909	1.1566
0.0095	622.21	5600	6.8088	1.1102
0.0077	666.63	6000	7.4451	1.1311
0.0062	711.11	6400	6.8486	1.0777
0.0051	755.53	6800	6.8812	1.1241
0.0051	799.95	7200	6.9987	1.1450
0.0041	844.42	7600	7.3048	1.1323
0.0044	888.84	8000	6.6644	1.1125
0.0031	933.32	8400	6.6298	1.1148
0.0027	977.74	8800	6.7860	1.1067

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.14.0
• Tokenizers 0.10.3