asr-africa/w2v-bert-2.0-yogera-lg-10hrs-v1

wav2vec2-bert

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the Yogera dataset. It achieves the following results on the evaluation set:

• Loss: 0.2858
• Wer: 0.1597
• Cer: 0.0355

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.8824	1.0	198	0.2803	0.2968	0.0591
0.2156	2.0	396	0.2128	0.2389	0.0493
0.1589	3.0	594	0.2110	0.2207	0.0458
0.1277	4.0	792	0.1942	0.1964	0.0422
0.1055	5.0	990	0.1698	0.1873	0.0390
0.087	6.0	1188	0.1771	0.1879	0.0428
0.0738	7.0	1386	0.1850	0.1856	0.0406
0.0589	8.0	1584	0.1799	0.1681	0.0381
0.0573	9.0	1782	0.1882	0.1863	0.0400
0.0481	10.0	1980	0.2275	0.1664	0.0359
0.0425	11.0	2178	0.2135	0.1696	0.0379
0.039	12.0	2376	0.2035	0.1600	0.0354
0.0351	13.0	2574	0.2095	0.1683	0.0366
0.0326	14.0	2772	0.2070	0.1589	0.0353
0.0302	15.0	2970	0.2526	0.1708	0.0367
0.0308	16.0	3168	0.2441	0.1642	0.0367
0.0255	17.0	3366	0.2504	0.1678	0.0365
0.0213	18.0	3564	0.2844	0.1721	0.0377
0.0225	19.0	3762	0.2602	0.1721	0.0383
0.02	20.0	3960	0.2746	0.1610	0.0351
0.0181	21.0	4158	0.2767	0.1668	0.0364
0.0149	22.0	4356	0.2442	0.1633	0.0355
0.0136	23.0	4554	0.2765	0.1677	0.0362
0.0156	24.0	4752	0.2858	0.1597	0.0355

Framework versions

• Transformers 4.45.2
• Pytorch 2.1.0+cu118
• Datasets 3.0.1
• Tokenizers 0.20.1