wav2vec2-xls-r-1b-ro

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - RO dataset. It achieves the following results on the evaluation set:

• Loss: 0.1113
• Wer: 0.4770
• Cer: 0.0306

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: 3e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.7844	1.67	1500	0.3412	0.8600	0.0940
0.7272	3.34	3000	0.1926	0.6409	0.0527
0.6924	5.02	4500	0.1413	0.5722	0.0401
0.6327	6.69	6000	0.1252	0.5366	0.0371
0.6363	8.36	7500	0.1235	0.5741	0.0389
0.6238	10.03	9000	0.1180	0.5542	0.0362
0.6018	11.71	10500	0.1192	0.5694	0.0369
0.583	13.38	12000	0.1216	0.5772	0.0385
0.5643	15.05	13500	0.1195	0.5419	0.0371
0.5399	16.72	15000	0.1240	0.5224	0.0370
0.5529	18.39	16500	0.1174	0.5555	0.0367
0.5246	20.07	18000	0.1097	0.5047	0.0339
0.4936	21.74	19500	0.1225	0.5189	0.0382
0.4629	23.41	21000	0.1142	0.5047	0.0344
0.4463	25.08	22500	0.1168	0.4887	0.0339
0.4671	26.76	24000	0.1119	0.5073	0.0338
0.4359	28.43	25500	0.1206	0.5479	0.0363
0.4225	30.1	27000	0.1122	0.5170	0.0345
0.4038	31.77	28500	0.1159	0.5032	0.0343
0.4271	33.44	30000	0.1116	0.5126	0.0339
0.3867	35.12	31500	0.1101	0.4937	0.0327
0.3674	36.79	33000	0.1142	0.4940	0.0330
0.3607	38.46	34500	0.1106	0.5145	0.0327
0.3651	40.13	36000	0.1172	0.4921	0.0317
0.3268	41.81	37500	0.1093	0.4830	0.0310
0.3345	43.48	39000	0.1131	0.4760	0.0314
0.3236	45.15	40500	0.1132	0.4864	0.0317
0.312	46.82	42000	0.1124	0.4861	0.0315
0.3106	48.49	43500	0.1116	0.4745	0.0306

Framework versions

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