xlsr-he-adap-es

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 2.3575
• Wer: 0.7473
• Cer: 0.3212

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: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
4.9752	3.3333	100	5.0053	1.0	1.0
3.5082	6.6667	200	3.7745	1.0	1.0
3.3178	10.0	300	3.4926	1.0	1.0
3.2202	13.3333	400	3.4051	1.0	1.0
3.2765	16.6667	500	3.4608	1.0	0.9692
3.1366	20.0	600	3.4355	0.9990	0.9359
3.1141	23.3333	700	3.3969	1.0	0.9301
2.6831	26.6667	800	2.8457	1.0239	0.7439
1.2342	30.0	900	1.7356	0.8983	0.4213
0.6381	33.3333	1000	1.3967	0.8541	0.3920
0.7105	36.6667	1100	1.4446	0.8261	0.3728
0.5292	40.0	1200	1.7486	0.8180	0.3681
0.2724	43.3333	1300	1.6140	0.8231	0.3571
0.3193	46.6667	1400	1.7165	0.8327	0.3538
0.304	50.0	1500	1.8261	0.7911	0.3372
0.1434	53.3333	1600	1.8850	0.7875	0.3361
0.151	56.6667	1700	1.9615	0.7824	0.3377
0.1663	60.0	1800	2.0116	0.7850	0.3352
0.1473	63.3333	1900	2.0547	0.7748	0.3309
0.1021	66.6667	2000	2.2370	0.7682	0.3305
0.1984	70.0	2100	2.2134	0.7560	0.3257
0.1731	73.3333	2200	2.2655	0.7483	0.3265
0.2154	76.6667	2300	2.3460	0.7514	0.3266
0.1106	80.0	2400	2.2466	0.7585	0.3260
0.1152	83.3333	2500	2.2369	0.7631	0.3309
0.0955	86.6667	2600	2.2950	0.7499	0.3223
0.1673	90.0	2700	2.3182	0.7453	0.3202
0.1635	93.3333	2800	2.3247	0.7428	0.3201
0.101	96.6667	2900	2.3576	0.7463	0.3215
0.1347	100.0	3000	2.3575	0.7473	0.3212

Framework versions

• Transformers 4.42.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1