asr-africa/whisper_BIG-C_BEMBA_1hr_v1

Whisper Small Bemba - Beijuka Bruno

This model is a fine-tuned version of openai/whisper-small on the BEMBA dataset. It achieves the following results on the evaluation set:

• Loss: 1.3624
• Wer: 0.6408
• Cer: 0.1701

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: 1e-05
• train_batch_size: 16
• 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: 500
• num_epochs: 100

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
2.6721	1.0	32	2.4561	1.5330	0.6440
2.2935	2.0	64	1.9492	0.9650	0.3391
1.7575	3.0	96	1.5668	0.8598	0.2686
1.4295	4.0	128	1.3788	0.7765	0.2352
1.1759	5.0	160	1.2559	0.7586	0.2371
0.9792	6.0	192	1.1833	0.7439	0.2093
0.8037	7.0	224	1.1540	0.6931	0.2193
0.6423	8.0	256	1.1380	0.6857	0.1988
0.462	9.0	288	1.1553	0.6905	0.2055
0.3141	10.0	320	1.1998	0.7169	0.2213
0.2089	11.0	352	1.2249	0.6661	0.1869
0.1359	12.0	384	1.2399	0.7071	0.2051
0.0831	13.0	416	1.2533	0.7148	0.2138
0.0553	14.0	448	1.2758	0.6553	0.1773
0.0334	15.0	480	1.3205	0.6406	0.1766
0.0286	16.0	512	1.2915	0.6486	0.1774
0.0248	17.0	544	1.3042	0.6628	0.1861
0.0177	18.0	576	1.3364	0.6628	0.1930
0.0131	19.0	608	1.3429	0.6505	0.1832
0.0116	20.0	640	1.3442	0.6763	0.2078
0.0105	21.0	672	1.3807	0.6548	0.1787
0.0056	22.0	704	1.3839	0.6480	0.1775
0.005	23.0	736	1.3965	0.6402	0.1775
0.0059	24.0	768	1.3936	0.6605	0.1941
0.0055	25.0	800	1.3850	0.6471	0.1871

Framework versions

• Transformers 4.44.2
• Pytorch 2.1.0+cu118
• Datasets 3.0.0
• Tokenizers 0.19.1