wav2vec2-large-xls-r-300m-hi-colab

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_13_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.7522
• Wer: 1.0223

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: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.9636	0.95	400	2.2746	1.0338
0.9533	1.89	800	0.8338	1.0699
0.5378	2.84	1200	0.7781	1.0159
0.4041	3.79	1600	0.7267	1.0191
0.3308	4.73	2000	0.6780	1.0237
0.2728	5.68	2400	0.6862	1.0203
0.2272	6.63	2800	0.6658	1.0266
0.198	7.57	3200	0.6819	1.0306
0.1787	8.52	3600	0.6930	1.0257
0.158	9.47	4000	0.7278	1.0318
0.1391	10.41	4400	0.7102	1.0319
0.1249	11.36	4800	0.7726	1.0190
0.1131	12.31	5200	0.7325	1.0253
0.1049	13.25	5600	0.7512	1.0227
0.095	14.2	6000	0.7580	1.0222
0.0835	15.15	6400	0.7161	1.0204
0.0817	16.09	6800	0.7530	1.0239
0.0707	17.04	7200	0.7613	1.0250
0.0682	17.99	7600	0.7412	1.0196
0.0599	18.93	8000	0.7700	1.0214
0.0593	19.88	8400	0.7522	1.0223

Framework versions

• Transformers 4.32.1
• Pytorch 2.2.0+cu121
• Datasets 2.12.0
• Tokenizers 0.13.2