arampacha/wav2vec2-xls-r-1b-hy-cv

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - HY-AM dataset. It achieves the following results on the evaluation set:

• Loss: 0.4521
• Wer: 0.5141
• Cer: 0.1100
• Wer+LM: 0.2756
• Cer+LM: 0.0866

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: 8e-05
• train_batch_size: 16
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
• lr_scheduler_type: tristage
• lr_scheduler_ratios: [0.1, 0.4, 0.5]
• training_steps: 1400
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
6.1298	19.87	100	3.1204	1.0	1.0
2.7269	39.87	200	0.6200	0.7592	0.1755
1.4643	59.87	300	0.4796	0.5921	0.1277
1.1242	79.87	400	0.4637	0.5359	0.1145
0.9592	99.87	500	0.4521	0.5141	0.1100
0.8704	119.87	600	0.4736	0.4914	0.1045
0.7908	139.87	700	0.5394	0.5250	0.1124
0.7049	159.87	800	0.4822	0.4754	0.0985
0.6299	179.87	900	0.4890	0.4809	0.1028
0.5832	199.87	1000	0.5233	0.4813	0.1028
0.5145	219.87	1100	0.5350	0.4781	0.0994
0.4604	239.87	1200	0.5223	0.4715	0.0984
0.4226	259.87	1300	0.5167	0.4625	0.0953
0.3946	279.87	1400	0.5248	0.4614	0.0950

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0