asr-africa/W2V2_Bert_DigitalUmuganda_Afrivoice_Shona_86hr_v1

facebook/w2v-bert-2.0

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

• Loss: 0.3279
• Wer: 0.2852
• Cer: 0.0608

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: 4
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.025
• num_epochs: 100

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.6237	1.0	3770	0.2098	0.2662	0.0444
0.198	2.0	7540	0.2007	0.2578	0.0431
0.1894	3.0	11310	0.1866	0.2487	0.0414
0.1734	4.0	15080	0.1879	0.2471	0.0430
0.1616	5.0	18850	0.1895	0.2596	0.0430
0.1535	6.0	22620	0.1861	0.2449	0.0419
0.1464	7.0	26390	0.1742	0.2410	0.0394
0.1404	8.0	30160	0.1716	0.2285	0.0377
0.1351	9.0	33930	0.1749	0.2323	0.0385
0.1284	10.0	37700	0.1792	0.2358	0.0391
0.1242	11.0	41470	0.1780	0.2355	0.0395
0.1169	12.0	45240	0.1938	0.2311	0.0389
0.1106	13.0	49010	0.1808	0.2289	0.0378
0.1041	14.0	52780	0.1838	0.2280	0.0381
0.0982	15.0	56550	0.1970	0.2274	0.0380
0.0916	16.0	60320	0.1861	0.2275	0.0376
0.0838	17.0	64090	0.1960	0.2306	0.0386
0.0781	18.0	67860	0.2029	0.2294	0.0380

Framework versions

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