firstcolab3

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_11_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2756
• Wer: 0.6226
• Cer: 0.0535

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: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
7.187	0.75	1000	3.7705	1.0	1.0
2.0277	1.5	2000	0.6139	0.9202	0.1545
0.8368	2.24	3000	0.4351	0.8589	0.1147
0.6772	2.99	4000	0.3762	0.8200	0.0990
0.5702	3.74	5000	0.3434	0.7889	0.0891
0.5205	4.49	6000	0.3427	0.7726	0.0855
0.4773	5.24	7000	0.3073	0.7408	0.0767
0.4389	5.98	8000	0.2969	0.7421	0.0759
0.4069	6.73	9000	0.2884	0.7134	0.0711
0.3858	7.48	10000	0.2952	0.7066	0.0699
0.36	8.23	11000	0.2846	0.6902	0.0662
0.3517	8.98	12000	0.2729	0.6756	0.0638
0.3265	9.72	13000	0.2844	0.6756	0.0645
0.3127	10.47	14000	0.2769	0.6803	0.0640
0.3016	11.22	15000	0.2772	0.6566	0.0618
0.2855	11.97	16000	0.2791	0.6540	0.0598
0.2699	12.72	17000	0.2714	0.6455	0.0589
0.264	13.46	18000	0.2782	0.6472	0.0588
0.2518	14.21	19000	0.2693	0.6398	0.0578
0.2498	14.96	20000	0.2761	0.6300	0.0561
0.2426	15.71	21000	0.2796	0.6366	0.0561
0.2271	16.45	22000	0.2804	0.6336	0.0554
0.2271	17.2	23000	0.2758	0.6347	0.0552
0.22	17.95	24000	0.2785	0.6279	0.0544
0.2143	18.7	25000	0.2783	0.6246	0.0538
0.2134	19.45	26000	0.2756	0.6226	0.0535

Framework versions

• Transformers 4.24.0
• Pytorch 1.12.1+cu113
• Datasets 2.7.1
• Tokenizers 0.13.2