wav2vec2-1b-E10_speed

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.8527
• Cer: 21.5519

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: 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: 50
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
21.0525	0.2580	200	4.6684	100.0
3.9541	0.5160	400	2.8257	57.0724
1.7271	0.7741	600	1.9449	44.8015
1.3002	1.0321	800	1.4433	34.1224
1.0165	1.2901	1000	1.3209	30.7272
0.895	1.5481	1200	1.2042	29.3351
0.8431	1.8062	1400	1.1021	27.5905
0.7042	2.0642	1600	1.3478	32.1311
0.6066	2.3222	1800	1.3199	32.1487
0.589	2.5802	2000	1.1196	27.7961
0.549	2.8383	2200	1.1813	29.9048
0.4874	3.0963	2400	0.9759	24.7415
0.3897	3.3543	2600	0.9338	23.9720
0.3993	3.6123	2800	0.9023	22.8560
0.361	3.8703	3000	0.8922	22.9911
0.3091	4.1284	3200	0.9250	23.5491
0.2812	4.3864	3400	0.8751	22.0982
0.2639	4.6444	3600	0.8540	21.5754
0.2451	4.9024	3800	0.8527	21.5519

Framework versions

• Transformers 4.45.2
• Pytorch 2.3.1.post100
• Datasets 2.19.1
• Tokenizers 0.20.1